# Andrew Leng Ly > Andrew Leng Ly is a designer, developer and analyst in Sacramento, CA. CSU Chico MIS graduate and founder of Parabox Digital, a web design studio. Open to Parabox Digital projects and work opportunities. Site: https://www.andrewlengly.com Also known as: Andrew Ly Short version of this file: https://www.andrewlengly.com/llms.txt ## Bio Andrew Leng Ly, Designer, Developer and Analyst. Based in Sacramento, CA. I'm Andrew. I studied Management Information Systems at CSU Chico and spent most of college leading student groups and running campus events. The part I liked best was taking a loose idea and turning it into a plan people can follow. It still is. Today I run Parabox Digital, a small studio that designs and builds websites for local businesses. When I'm offline I'm usually on my bike or reading about business and wellness. ## So, who's Andrew? ### Who are you and what do you do? I run Parabox Digital, a small web studio in the Sacramento area where I design and build websites for local businesses. I studied Management Information Systems at Sierra College and CSU Chico as a first-generation college student. Along the way I was a student advocate and competed in hackathons and DataFest. Big fan of learning through life! ### How'd you get into it? My dad ran a computer business, and I started out building Windows Vista era computers for him. That got me into hardware first, then networks, coding and UX design. ### How do you see yourself? A learner. A doer. A bit optimistic. A bit curious. A bit passionate. A little extraverted and introspective at the same time. Honestly, the more I learn about myself, the less I can put a label on it. ### Has that changed over time? Absolutely. Growing up and through college, I was almost the opposite of who I am now. A lot of that changed when I started learning out of passion. Before, I learned for outcomes. With that came a love for learning from other perspectives. ### Who inspires you? My girlfriend! My sisters too. Then the mentors and professors who guided me through college, and the peers who backed me in Delta Sigma Pi, BIS Society, Associated Students of Sierra College and Chico State, and plenty more. People I haven't met yet: Charlie Munger, James Dyson, Steve Wozniak and Akio Morita. ### What's Parabox Digital? It's the web studio I founded and run, out of passion for design, development and business. Past the website itself, I handle SEO and ongoing upkeep. My clients own their sites, and I stay on to keep them running well. Working with clients and seeing the work make a real difference for their business is what makes it meaningful to me. ### Why'd you start it? A capstone project turned into a real client, and I kept growing it from there to support businesses my own way. Running it as a professional service lets me work with more businesses and keep learning new tech as the industry moves. ### What's on your playlist? Playlist, or playlists? 😏 andrewlengly.com/music ### What do you put on when you need to lock in? Ambient sounds for designing and building. Something more intense for working out. Driving the wavy mountain cliffs? Nothing 😭 ### Is there a song you'll never skip? Sometimes I skip a song just to hit the skip button. ### Where'd you grow up? Rio Linda, a small town just outside Sacramento with a country feel. We had a small house on a multi-acre yard. As a kid I was outside playing with sticks, mud, chickens and frogs, then back inside learning tech. It was a blast! ### How did it shape you? Growing up with both kept my mind open to new experiences, and to learning by studying and by doing. Once both count as learning, you can learn while playing. If play is a form of learning, and so is work, then work can be play. I try to carry that idea with me every day. ### What do you enjoy doing? Outside of learning, Zankar and I love exploring new cafes and libraries. New gyms too. Outdoors, I look for trails to walk or bike. I'm most creative when I'm moving. Indoors, it's reading, working out and cozy games. ### How'd you get into that? Exploring with Zankar is a great way to get to know our area, and getting outside balances all the time I spend inside working. A new gym is its own little adventure. What machines do they have? Is there a pool, a sauna, outdoor training? Mostly I'm looking for ways to stay healthy, physically and mentally. ### What's your max bench? I've never maxed out. Last time I checked, 275 for 3 reps. ### What are you reading right now? The Intelligent Investor by Benjamin Graham, Deep Work by Cal Newport and Zero to One by Peter Thiel. Got recommendations? Reach out! ### What video games do you play? Right now it's Stardew Valley, Cyberpunk 2077 and MapleStory. ### What do you want to try next? Skydiving! ### What are you building right now? Outside of websites, two things. One is an app for building social, active third spaces. The others are apps focused on safety. This one is always changing, so reach out for the latest! ### What are some ways to contact you? LinkedIn is the best way. linkedin.com/in/andrew-leng-ly For Parabox Digital, email andrew@paraboxdigital.com or book time at paraboxdigital.com/book. Or email me at andrew.leng.ly@gmail.com! ### What should someone reach out about? Websites, tech, questions, or if you're a recruiter looking to hire ☺️ ## Case studies ### Grid Watch case study Page: https://www.andrewlengly.com/grid-watch-casestudy Open Grid Watch: https://www.andrewlengly.com/gridwatch Read the methodology: https://www.andrewlengly.com/gridwatch/methodology Code: https://github.com/AndrewLengLy/gridwatch (Private) Built with: Next.js, MapLibre, deck.gl, Supabase, PostGIS, Open-Meteo, Claude Code Grid Watch is a live map of California power outages set over fire weather, built because the state's outage feed shows only what is out right now. It saves a snapshot every 15 minutes and lines each outage up with the Red Flag Warnings around it. It is built with Next.js, MapLibre, deck.gl and Supabase with PostGIS, and it has been live since November 5, 2026. - The problem: California’s outage feed only shows what’s out right now. Once the lights come back, that outage is gone from the record. - What I did: Built a map that saves the feed every 15 minutes and lines each outage up with the Red Flag Warnings around it. - Where it stands: Live since November 5, 2026. The record grows with every snapshot, and the first big event will get its own page. #### Why I built it. When the lights come back, the outage disappears. California utilities cut power on purpose when wind and dry brush make a downed line dangerous. In October 2019 those shutoffs first reached about 800,000 customer accounts, roughly 2.5 million people. Cal OES runs a live statewide outage map, and it only shows what’s out right now. Once power returns, the outage drops off the feed. Nobody can go back and check how long a county stayed dark, or whether the outage sat inside a Red Flag Warning. So I built the record myself. Every 15 minutes Grid Watch saves the feed along with the alerts in force at that moment. Source: 2019 California power shutoffs, https://en.wikipedia.org/wiki/2019_California_power_shutoff Who it’s for: - Residents: Checking whether their ZIP is out and what the weather is doing nearby. - Reporters: Who need a customer count with a timestamp and a named source mid-event. - Local officials: Comparing this event with the last one in their county. - Researchers: Testing whether shutoffs line up with the weather the utilities cite. #### What it does. A live map with a memory. - Live map: Outages sized by customers out, over NWS fire and wind alerts and active fire perimeters, on 3D terrain. - 48-hour replay: Drag the scrubber to play back each snapshot and watch an event spread and restore. - Overlap readout: The share of customers out inside a Red Flag Warning, next to a 30-day baseline for the same zones. - My area: Type a ZIP or an address to see the outages and warnings within 10 miles. - Event pages: Any statewide event past 50,000 customers out gets a page with its peak and restoration curve. - County pages: One page for each of the 58 counties, with current status and history. - Methodology: Every formula and threshold in plain language, linked to its source. #### How I built it. When a live feed disagreed with the spec, the feed won. From the live feeds to the first scheduled snapshot: 1. Every source, live. Cal OES, NWS, NIFC, map tiles and geocoders. 14 of 14 answered in CI on November 5. Caught: A module error that hid the check on its first CI run and would have broken the backup job 2. Rules from the replies. Paging at the real cap of 1,000 records. The misspelled county field, read as spelled. Planned: Covers maintenance and shutoffs alike, so cause text splits the two 3. Pipeline tests. 38 Vitest tests, 7 on real Postgres with PostGIS. 8 Playwright passes at 375 and 1440 px 4. Launch checks. Four pages scored by Lighthouse on mobile. 90 files through the copy check. Color: The map palette measured under simulated deuteranopia 5. Live on Nov 5. pg_cron took the first scheduled snapshot at 1:45 PM #### The questions. Each view answers one question. - How many customers are out right now, and where? - Metric: Customers out, summed by county and utility - Shows on: Live map, county pages - Which outages are fire shutoffs, and which are maintenance? - Metric: Shutoff class from the cause text - Shows on: Map legend and counts - What share of customers out sit inside a Red Flag Warning? - Metric: Inside share - Shows on: Overlap readout - Is that share higher than on calm days in the same zones? - Metric: Baseline share and lift - Shows on: Overlap readout - How long do outages last? - Metric: Median and longest duration - Shows on: Live stats, county pages - When does a bad day count as an event? - Metric: Open and close thresholds - Shows on: Event pages - How fast does power come back after the peak? - Metric: Hours to 50% and 90% restored - Shows on: Event pages #### How it works. A scheduled job writes the record every 15 minutes. How a snapshot moves from the feeds to the page: 1. Cal OES and NWS. Outage points and areas. Fire and wind alerts 2. Snapshot job. Supabase Edge Function. Reconciles open outages. One transaction per run. Runs on: pg_cron every 15 minutes, with a GitHub Action as backup if data goes 30 minutes stale 3. Postgres + PostGIS. Tables closed by row-level security. Public reads through views. 90 days of raw rows 4. Next.js site. Served at /gridwatch. Cached server pages. Map API routes. Also reads: NIFC fire perimeters, live and cached for 15 minutes An outage opens the first time its incident ID shows up. It closes once it’s missing from 2 successful snapshots in a row. A failed fetch never counts as a miss. That way a feed hiccup can’t make it look like power came back. #### The math. Every number on the map has a rule behind it. Counts are customer accounts. One account is one meter, which can cover a whole household. Customers out and duration: - customers out = the sum of customers across open incidents in a snapshot - duration = last seen minus start The start is the utility’s own start time when it’s no later than Grid Watch’s first sighting. A later utility time is treated as a data error, and the first sighting wins. Shutoff or maintenance: - shutoff = a planned outage whose cause text names a shutoff or fire risk Cal OES files routine maintenance as Planned too. Terms such as PSPS, power shutoff, wildfire and fire weather mark the real shutoffs. Every other planned outage counts as maintenance. How the overlap readout compares a warning with calm days Fire-weather overlap: - inside share = customers out inside active Red Flag Warnings / all customers out - baseline share = customers out inside the same zones on calm snapshots / all customers out on those snapshots - lift = inside share / baseline share Calm snapshots come from the last 30 days with no Red Flag Warning anywhere in California. Shutoffs and unplanned outages get scored on their own, since utilities schedule shutoffs because of the warning. The site calls lift a correlation. Events and restoration: - event opens = when statewide customers out reaches 50,000 - event closes = after customers out stays under 10,000 for 6 hours - time to 50% = hours from the peak to the first snapshot at or under half the peak - time to 90% = hours from the peak to the first snapshot at or under a tenth of the peak Thresholds live in a config table. Restoration times carry 15 minutes of resolution, since that’s how often snapshots land. #### What it found. The first Red Flag Warning on the record came and went with no shutoff reported. - 802 Snapshots stored - 209 Snapshots with a Red Flag Warning up - 0 Snapshots with a shutoff - 0 Failed Cal OES fetches Peak statewide customers out each day, as of 9:45 PM on November 13, 2026: - Nov 5 (from 1:36 PM): 9,027 - Nov 6: 10,097 - Nov 7: 7,049 - Nov 8: 8,118 - Nov 9: 12,642 - Nov 10: 13,415 - Nov 11: 8,015 - Nov 12: 15,684 - Nov 13 (to 9:45 PM): 16,728 Days run midnight to midnight, Pacific time. A full bar would be 50,000, the count that opens an event. NWS Los Angeles/Oxnard issued a Red Flag Warning at 1:36 PM on November 5, and the first scheduled snapshot caught it nine minutes later. It stayed up until 6 PM on November 7. No planned outage named a shutoff or fire risk as its cause, then or since. The highest count so far was 16,728 customers out, at 1:30 PM on November 13. An event opens at 50,000, so the events page is still empty. Since its first run, pg_cron hasn’t missed a quarter hour. These figures run through the 9:45 PM snapshot on November 13, 2026. Source: Grid Watch events, https://www.andrewlengly.com/gridwatch/events #### What the data got wrong. The raw feeds broke a few of my assumptions. I pulled every feed live before writing a single metric. Each of these would have changed a number on the map. - “Planned” includes routine maintenance - Risk: Shutoff totals inflated by scheduled work - Call: Sort shutoffs by cause text - Utilities schedule shutoffs because of the warning - Risk: Shutoffs inflate the overlap for unplanned outages - Call: Score the two kinds on their own - The customers field is spelled Impacted_Cutomers_Not_Planned - Risk: A corrected spelling reads zero customers with no error - Call: Read the field as published - The service hands back 1,000 records at a time - Risk: A big event gets undercounted with no error - Call: Page through every record each snapshot - Only some utilities publish outage areas - Risk: Area-based totals skip whole utilities - Call: Count from incident points and draw areas for context - A failed fetch looks like zero outages - Risk: Outages close early and restoration looks faster - Call: A failed fetch never counts as a miss - Many NWS alerts list zone IDs with no outline - Risk: Warnings missing from the map and the overlap - Call: Fetch each zone’s outline, cache it, and refetch when a new ID appears - Calm snapshots have no warning zones of their own - Risk: The baseline reads 0% and lift breaks - Call: Score calm snapshots against the current warning’s zones #### Decisions. Built so a bad feed can’t rewrite history. - One transaction per snapshot: A crash mid-run leaves no half-written snapshot behind. - Close after 2 straight misses: One dropped row doesn’t end an outage early. - Metric rules as pure TypeScript: The same code runs in Deno on Supabase and in Node for the backup job and the tests. - Tests against real Postgres: Spatial joins and the transaction get tested the way they run in production. - Reads only through views: Every table is closed by row-level security, so the public site can read the record and has no way to change it. - 14 live endpoints checked in CI: Every pull request calls each source once and writes its status and record count to the run summary. - A backup schedule: GitHub Actions takes a snapshot whenever data goes 30 minutes stale. - MapLibre pinned to 5: Version 6 broke the deck.gl terrain overlay and crashed the map on load. #### Design. The map reads like an instrument panel at night. Outages glow amber over a near-black map, and Red Flag Warnings get red hatching. Every color on the map also has a shape, so nothing depends on color alone. Grid Watch’s map key, color and shape together - Type: Chakra Petch for headings, Inter Tight for text, and Geist Mono for every number so digits hold still as counts change. - Phones first: At 375 px the panels sit in a bottom sheet you can drag, and the map keeps the whole screen. - Reduced motion: The event replay doesn’t autoplay, and the scrubber steps between frames. Color distance between map colors (ΔE), normal vision and simulated deuteranopia (Normal, Deuteranopia): - Outage vs wind alert: 113.7, 112.5 - Outage vs Red Flag Warning: 67.4, 40.5 - Red Flag Warning vs fire perimeter: 66.3, 33.8 - Outage vs fire perimeter: 55.8, 52.2 - Outage vs planned maintenance: 33.3, 31.3 A script simulates deuteranopia and measures each pair that shares the map. The closest pairs also differ by shape, such as hatching or a ring. #### The 3D map. In 3D you can see which way the wind blows over each outage. I modeled this view on FireTrace, Dilum Sanjaya’s wildfire map. It recolors its satellite imagery with a Mapbox paint property that MapLibre doesn’t have. So Grid Watch repaints each USGS imagery tile in the browser before the map draws it. A custom tile protocol maps every pixel onto a warm charcoal ramp, and water stays dark. Source: FireTrace, https://firemap-two.vercel.app/ - Outage columns: From zoom 8.5 each outage stands as a column. Shutoffs show as wireframes, and the five tallest carry a chip like PG&E PSPS · 12,400 out · since 6:10 AM. - Wind: Streaks drape over the terrain from an Open-Meteo forecast grid, colored by gust. With reduced motion on, they hold still as arrows. - Transmission lines: The California Energy Commission’s map of overhead lines at 60 kV and up. Pieces inside a Red Flag Warning glow red at the frame you’re on. - Conditions card: Click the map for wind, gusts, humidity and temperature at that hour, plus the elevation and the nearest outage. - Fly in downwind: Click an outage of 1,000 or more customers and the camera flies in close, facing the way the wind blows. - Key events: Shutoff starts, warnings going up and coming down, the peak and the biggest restore. Each row jumps the replay to that moment. - Event view: Event pages open low over the terrain at 2x height, with the same controls as the live map. Column height: - height = 18 m × √customers out, capped at 4,500 m A square root keeps big shutoffs readable next to small outages. 100 customers stand 180 m tall and 10,000 stand 1,800 m. The cap arrives at 62,500. Wind and lines: - state grid = about 170 points at 0.5°, cached 3 hours - close-up grid = up to 300 points at about 0.1° around a zoomed 3D view, cached 3 hours - line in a warning = a piece of line 2 km or shorter whose midpoint sits inside an active Red Flag Warning Open-Meteo’s free tier allows 10,000 calls a day, and each point counts as one. Close-up grids snap to half degrees so nearby views share a cache, which keeps the worst case near 1,400 calls for the state grid plus 2,400 per busy view. #### Where it stands. Live, with the record still filling in. - 15 min Between snapshots - 58 County pages - 31 Unit tests passing, plus 7 that run on real Postgres - 14/14 Live source checks passing in CI The first source check saw 263 outage incidents and 121 outage areas. No fire or wind alerts were up that morning, and a Red Flag Warning followed in the afternoon. Lighthouse on mobile, run against sample data (Performance, Accessibility, Best practices, SEO): - County page: 96, 100, 100, 100 - Methodology: 96, 100, 100, 100 - Live map: 69, 100, 96, 100 - Event page: 68, 100, 96, 100 The map pages are the weak spot. The home page paints in 1.0 s with zero layout shift, then spends 4,100 ms of blocking time setting up the WebGL map and terrain. #### What’s next. What I’d do with a full fire season of data. Next: - The first big event: Write it up once something passes 50,000 customers, with its restoration times and lift. - Shutoffs outside warnings: Measure how many shutoff customers sat outside any Red Flag Warning, by utility. - A faster map: Load terrain and the outage layer after first paint to cut the 4,100 ms of blocking time. - Open data: Publish a daily CSV of statewide totals so others can check the work. Known limits: - Reported data: Grid Watch shows what utilities report to Cal OES. Each utility’s own map stays the official word, and the site links to it. - Correlation: Overlap can’t show that a warning caused a shutoff. - Shutoff labels: Grid Watch spots a shutoff by its cause text. One filed as Planned under a vaguer cause counts as maintenance. - Line status: The transmission layer shows where lines run. It can’t say whether a line is energized. - Modeled wind: Wind is a model forecast for 10 m above the ground. Replay frames older than 48 hours show none. - A young record: History starts in November 2026, so the 30-day baseline is still filling in. #### TL;DR - Built: A live map of California power outages over fire weather, with a snapshot saved every 15 minutes. - Found: The first Red Flag Warning on the record came and went with no shutoff reported. - By the numbers: 802 Snapshots stored | 209 Snapshots with a Red Flag Warning up | 0 Snapshots with a shutoff - The catch: The record is young. One quiet warning says little about a full fire season. - Next: Write up the first big event once statewide customers out pass 50,000, with its restoration times. ### Chain Routes case study Page: https://www.andrewlengly.com/chain-routes-casestudy Open Chain Routes: https://chains.andrewlengly.com Read the methodology: https://chains.andrewlengly.com/methodology Code: https://github.com/AndrewLengLy/chainline (Private) Built with: Next.js, MapLibre, deck.gl, d3, Supabase, GitHub Actions, Claude Code Chain Routes is a 3D terrain map of Sierra chain controls that saves every checkpoint every 10 minutes, since Caltrans publishes the current status only. Each storm between Sacramento and Tahoe can be replayed. It is built with Next.js, MapLibre, deck.gl and Supabase, and the snapshots it has stored since November 7, 2026 go public once the first build merges. - The problem: Caltrans publishes chain controls as current status only. Nobody can look up what I-80 did last Saturday. - What I did: Built a map that saves every checkpoint every 10 minutes and lines it up with snowfall and the next storm at each summit. - Where it stands: Tested on sample storms, with snapshots stored since November 7. That history goes public once the first build merges. #### Why I built it. Chain controls drop out of the record the moment they lift. Chain Routes answers one question. Should I drive up this weekend, and when? Caltrans publishes chain control status as current-only. Once controls lift, they’re gone from the feed, so nobody can check how long Donner sat at R-2 last Saturday or how fast it cleared. So Chain Routes keeps its own record. Every 10 minutes it reads every checkpoint between Sacramento and Tahoe and stores each change. Source: Caltrans CWWP2, https://cwwp2.dot.ca.gov #### What it does. The passes and Tahoe share one map. - Passes map: Chain controls and closures on the Sierra passes from Sacramento to Tahoe, over snowfall on 3D terrain. - Tahoe Basin map: The lake loop and its summits, with Nevada 511 conditions placed on the same R-level ramp. - Replay: A scrubber that steps through each 10-minute frame. Watch a storm close the passes and clear them. - Trip check: Pick one of 16 routes and a departure up to 72 hours out. It estimates when you reach each summit and shows the NWS forecast for that hour. - Summit pages: 13 pinned summits, each with current conditions and its last five storms. - Storm log: Each storm gets a page with its hours under control by route and its time to clear at each summit. - Methodology: Every source and formula in plain language, plus the full chart of status codes. #### How I built it. I started from what the feeds really send. From a live source check to the 10-minute schedule: 1. Source check. One script calls every source once. CI runs it on every pull request 2. Test fixtures. Trimmed copies of live responses. Nevada 511 samples written from its docs 3. Tests. Unit tests on saved responses and SQL migrations. 20 of 20 Playwright tests on sample storms. Crawl: 39 links, none broken 4. Color review. Deuteranopia simulated on each status step. I kept R-2 and R-3, the closest pair at ΔE 13.3 5. Schedule. The snapshot job started November 7. Runs: 852 by November 12, with no failed jobs Where a feed broke one of my assumptions, the parser or the rule changed to fit the feed. Source: What the data got wrong, https://www.andrewlengly.com/chain-routes-casestudy#data #### The 3D map. The snow level sits right on the mountains. I modeled the map on FireTrace, Dilum Sanjaya’s wildfire map. It recolors its imagery with a Mapbox paint property that MapLibre doesn’t have. So Chain Routes repaints each USGS imagery tile in the browser onto a night blue-gray ramp. The ramp stays dark so every route keeps its contrast. Source: FireTrace, https://firemap-two.vercel.app/ - Snow level: The NWS snow level at each summit, spread across the map. Ground above it turns frosty, under a labeled ice-blue line. - Follows your trip: The line uses the current hour, or your departure once a trip check is set. Snow level is a forecast and isn’t stored, so past replay frames leave it off. - Conifers: Cones on OpenStreetMap woodland at three times life size, mostly between 3,000 and 8,500 ft and thicker on north-facing slopes. Above the snow level they get white caps. - Checkpoint posts: Controlled checkpoints stand as posts in their status color. Up to six carry a chip like R-2 · since 4:10 AM, most severe first. - Key events: The first control on each route, closures and reopenings, each route’s peak and the all clear. Each row jumps the replay and flies to the checkpoint. - Terrain height: 1x, 1.5x, 2x or 3x, saved in the link so a shared view opens the same way. - snow level = sum of each summit’s level ÷ d² / sum of 1 ÷ d² Here d is the distance to each pinned summit in km, never under 0.5. Frost covers ground above that level on a 384 by 384 elevation grid from zoom 9 terrain tiles, or zoom 10 for the Tahoe Basin. Away from the summits the line is an estimate. #### How it works. A scheduled job keeps the record every 10 minutes. How a snapshot moves from the feeds to the page: 1. Caltrans and Nevada 511. Chain control status. Lane closures and road weather 2. Snapshot job. Supabase Edge Function. Stores each change. Checks every record’s time. Runs on: pg_cron every 10 minutes 3. Postgres. Its own chainline schema. Public reads through views. Shares a free project with Grid Watch 4. Next.js site. Passes and Tahoe maps. Storm and summit pages. Also reads: NOHRSC snowfall grids from a GitHub Actions job twice a day, plus NWS forecasts When the database is missing or fails, the site reads the feeds directly. It shows one frame with no past, and the map keeps working. #### The math. Each number has one fixed definition. Every day and time boundary uses Pacific time. Definitions stay fixed once published, so this season can be compared with the next. - hours(route, level) = time when the highest checkpoint on the route sits at that level A route counts while at least one checkpoint is above R-0. Overlapping checkpoints never double count. - storm opens = when a checkpoint on I-80, US-50, SR-89 or SR-267 leaves R-0 for 3 hours or more - spells merge = when they fall less than 12 hours apart - time to clear = hours from the last hour of snow depth gain over 0.5 in. to R-0 at every checkpoint near the summit A Nevada basin segment reporting chains or a closure for 3 hours also opens a storm. A spell longer than 14 days counts as a seasonal closure and can’t start one. - weekend odds = Saturday windows with chain controls / Saturdays this season with data A Saturday window runs 6:00 to 10:00 a.m. It counts as controlled if any checkpoint shows a status above R-0 in any frame inside it. I-80 is scored westbound and eastbound on their own. - threshold = median NOHRSC 24-hour snowfall at the summit for the period when controls went up The summit page waits for 10 storms before it shows one, and prints the sample size beside it. #### What the data got wrong. The live feeds broke a few of my assumptions. I pulled every feed live before writing a metric. Each of these would have written false history. - Caltrans sometimes serves an old copy of a file - Risk: Months-old readings written as fresh history - Call: Check record times, and mark a district down when nothing is newer than 2 hours - Some weather stations stay in service long after their last reading - Risk: Readings up to 2,270 days old on the map - Call: Drop readings older than 3 hours - District 3 lists each checkpoint several times with letter suffixes - Risk: One checkpoint counted three times - Call: Collapse matching rows into one - The District 10 weather file ships malformed JSON - Risk: A whole district’s road weather fails to parse - Call: Repair the missing commas before parsing - Nevada has no R codes - Risk: Two states on two scales - Call: Place each Nevada condition on the nearest step and print the Nevada wording beside it - Supabase returns at most 1,000 rows per request - Risk: Replay frames and season numbers cut off with no error - Call: Page through every long read - The Central Sierra Snow Lab reports no hourly depth - Risk: Time to clear at Donner drops to 12-hour resolution - Call: Read depth from Independence Camp and keep the lab’s water content #### Decisions. Built so a stale file can’t write false history. - A line for each direction: Eastbound and westbound have their own road lines, so a segment’s status follows the way you’re driving. - Retention that keeps the numbers true: Raw rows age out after 120 days. The rows that weekend odds and closure totals depend on stay. - Truck-only codes on the R-1 step: They restrict some traffic with no rule for cars. - A schema of its own: Chain Routes lives in a chainline schema inside the free Supabase project that runs Grid Watch. - Its own secret for the schedule: The cron call signs in with a random secret kept in Vault, so the service role key never goes there. - The map loads after the page: It starts 1.2 s after load on desktop and 2.5 s on a phone, or as soon as you interact. #### Design. Roads stay the brightest thing on the map. Chain levels run from yellow to red over a near-black map, and closed roads turn white with a hatch. Every status prints its Caltrans code and plain meaning beside the color. - Type: Overpass for text and Overpass Mono for every code and reading. - Feed down in gray: When a feed goes quiet, the map keeps the last known status in gray and prints when the feed went down. Color distance between neighboring status steps under simulated deuteranopia (ΔE) (ΔE, Lightness): - None to R-1: 90.3, 29 to 87 - R-1 to R-2: 23.1, 87 to 70 - R-2 to R-3: 13.3, 70 to 57 - R-3 to closed: 59.2, 57 to 100 Every pair stays at least 10 apart. R-2 and R-3 sit closest, and both print their code. #### Where it stands. Tested on sample storms and waiting on a real one. - 10 min Between snapshots - 13 Summit pages - 16 Routes in the trip check - 130 Unit tests passing The last source check ran 29 sources with no failures and skipped Nevada 511, which needs a key. Its parsers are built from the published docs. Lighthouse on mobile, run against sample data (Performance, Accessibility, Best practices, SEO): - Passes map: 100, 100, 100, 100 - Tahoe Basin map: 100, 100, 100, 100 - Donner Summit page: 100, 100, 100, 100 Read the map scores with care. Lighthouse stops before the map starts, so they measure the server-rendered page. With a laptop GPU at 4x CPU throttle, the map is ready in about 3.0 s at 375 px and 1.6 s at 1440 px. #### What’s next. What I still need to check. Next: - The first real storm: Swap the sample storm on this page for a real one once history builds. - A live Nevada check: Run the Nevada parsers against a real 511 key. - A phone test: Map timings so far come from a laptop. A real phone will differ. - The snowfall upload: Turn on the upload to Supabase storage. So far the grid job has run against NOHRSC with the upload off. Known limits: - Up to 10 minutes old: Chain rules change by the minute, so Caltrans QuickMap and Nevada 511 have the final word. - No road forecast: The forecast covers snow. It can’t say when Caltrans will put chains up. - Early odds: Weekend odds describe this season so far and carry little weight early on. #### TL;DR - Built: A 3D terrain map of Sierra chain controls from Sacramento to Tahoe, saved every 10 minutes. - Found: 130 unit tests pass, and the last source check ran 29 sources with no failures. - By the numbers: 130 Unit tests passing | 29 Sources in the last check, no failures | 16 Routes in the trip check - The catch: Every result so far comes from sample storms. The page is still waiting on a real one. - Next: Replace the sample storm with the first real storm once enough history builds up. ### Fish Maps case study Page: https://www.andrewlengly.com/fish-maps-casestudy Open Fish Maps: https://tides.andrewlengly.com Read the methodology: https://tides.andrewlengly.com/methodology Code: https://github.com/AndrewLengLy/tideline Built with: Next.js, MapLibre, deck.gl, d3, Claude Code Fish Maps is a 3D map of San Francisco Bay that shows how much sea level rise turns the year's highest tide into an ordinary high tide. It ranks the highest predicted tides at 7 Bay gauges and draws NOAA's sea level rise water on the shoreline one foot at a time. It is built with Next.js, MapLibre, deck.gl and d3, and it has been live since November 6, 2026. - The question: How much sea level rise turns today’s highest tide of the year into an ordinary high tide, and where does that water go? - What I did: Built a 3D map that ranks the year’s highest predicted tides at 7 Bay gauges and draws NOAA’s sea level rise water on the shoreline, one foot at a time. - Where it stands: Live since November 6, 2026. Every source works with no key, and there’s no database. #### Why I built it. A king tide is a preview of an ordinary high tide. King tides are the highest predicted tides of the year. In San Francisco Bay they arrive around the winter and summer solstices. On a calm day one may only wet low paths and parking lots. With enough sea level rise, that same water shows up on an ordinary day. Fish Maps puts a number on that gap at each gauge and draws the matching water on the shoreline. Source: California King Tides Project, https://www.coastal.ca.gov/kingtides/ #### What it does. Everything sits on one map of the Bay. - 3D map: Seven Bay gauges on 3D terrain, with NOAA’s mapped water for 0 to 10 ft of sea level rise. - Tide scrubber: Plays the predicted tide and adds it to the slider, so the shoreline moves with the water. - King tide calendar: The 20 highest predicted tides in the next 12 months across the Bay, from NOAA alone. - Check a place: Tap the map or search an address to see the first foot of rise where that spot goes under, and roughly when the state expects it. - Go see it: The next king tide, when to arrive, a warning for peaks after dark and spots to stand at near the gauge. - Station pages: One page per gauge, with predicted days over its flood level today and with 1, 2 and 3 ft of rise. - Shoreline spots: 15 places people go to watch king tides, each with the first NOAA step that floods it. #### How I built it. Checks changed the build right up to launch. From 87 source calls to the first real deploy: 1. 87 source calls. 86 came back OK. Skipped: USGS CoSMoS. Its map service returned the one 404 2. NOAA’s live answers. Thresholds taken from station metadata. Sea level rise tiles at zoom 16, about 2.4 m a pixel 3. Spots traced to sources. Every fact matched to a source page, then re-checked. Wording scaled to the evidence behind it. Moved: Sausalito, from Bridgeway to Kappas Marina, where the flooding is documented 4. Tests before launch. 53 unit tests, every derived metric covered. 25 browser tests at phone and desktop widths. Bug: Floating-point math pushed a tide at the threshold over it 5. First real deploy. Flat land color hid the poster image on a cold load. Until imagery arrived, about 3 s on fast wifi. Fixed: The canvas stays transparent until imagery loads, up to 8 s, and a browser test covers it Some NOAA services answered wrong. The map works around each one. Source: What the data got wrong, https://www.andrewlengly.com/fish-maps-casestudy#data How a spot page words its evidence: - water reported = a source reports water at the spot during king tides (8 spots) - viewing event = a source reports a viewing event, and the page says only that (4 spots) - dated photo = a dated photo caption is the only record (2 spots) - nothing found = no source documents king tide flooding there, and the page says so (Redwood Shores) At the Embarcadero, reports show water at king tides, and NOAA’s layer first floods it at 4 ft. Its page says so. That gap is a limit of the screening layer. #### The 3D map. Buildings turn coral when the water reaches them. I modeled the map on FireTrace, Dilum Sanjaya’s wildfire map. It recolors its imagery with a Mapbox paint property that MapLibre doesn’t have. So Fish Maps repaints each USGS imagery tile in the browser onto a gray-blue ramp. It keeps a quarter of the original color, so marsh and mudflat still read apart from pavement. Source: FireTrace, https://firemap-two.vercel.app/ - Buildings in the water: From zoom 15, every building on screen is tested against NOAA’s connected water at the foot the map shows. Wet ones turn coral, and the slider panel counts them. - Staff gauges: Each gauge stands as a staff marked every foot. A 6 ft figure stands at the high-tide line for scale. - Key events: Each day’s highest high this week, king tides, crossings of the minor flood level and the largest observed surge. Each row jumps the scrubber. - Terrain height: 1x, 2x, 3x or 5x, saved in the link. The Bay floor is set flat so the water stays one surface. - coral building = one whose footprint center sits in NOAA’s connected water at the whole foot the map shows That’s one point per building tested against a screening layer, so a coral building may stay dry and a gray one may not. Up to 1,500 buildings are tested at a time, or 600 on phones. #### How it works. No database, and no keys. How a tide moves from NOAA to the map: 1. NOAA CO-OPS. Tide predictions. Observed water levels. Datums and flood levels 2. Next.js server. Cached pages. King tide and flood-day math. Refreshes: Predictions once a day from a Vercel cron, observed water every 6 minutes 3. Browser map. MapLibre with 3D terrain. deck.gl gauge staffs. Also reads: NOAA sea level rise tiles, repainted pixel by pixel NOAA caps high and low requests at a year and 6-minute requests at a month, so longer ranges get split. Datums and flood levels change rarely, so a script builds them once and they’re committed. #### The math. Every number starts from mean higher high water. Mean higher high water (MHHW) is the average height of each day’s higher high tide. It’s where the shoreline sits on an ordinary day. - height above MHHW = predicted height minus MHHW - future-tide equivalent = the larger of 0 and the height above MHHW - surge = the largest gap between observed and predicted water in the last 72 hours The equivalent is the sea level rise at which a water level becomes an ordinary daily high. It assumes the tide range holds as the sea rises. - days(+x ft) = Pacific days with a predicted high where height + x minus MHHW tops NOAA’s minor flood level Counted for 0 to 3 ft of rise over the next 12 months. Three gauges have only an action level, and their pages count that and say so. - level = slider + the tide above MHHW at the selected gauge - first flood step = the lowest foot where a quarter of a 5 by 5 grid of points at the spot is wet and connected NOAA publishes one layer per whole foot. At 2.4 ft the map draws the 2 ft water plus 40 percent of the band between 2 and 3 ft. That blend is visual, since NOAA models no fractional feet. #### What it found. At Alameda, 2 ft of rise tops the flood level on more than half the days of the year. Predicted days over the minor flood level at Alameda, as of November 6, 2026: - No rise added: 0 - 1 ft of rise: 22 - 2 ft of rise: 198 - 3 ft of rise: 341 Each bar adds its rise to every predicted high in the next 12 months. A full bar would be all 365 days. Predicted days over each gauge’s threshold with 2 ft of rise, as of November 13, 2026: - Point Reyes: 226 - Port Chicago: 219 - San Francisco: 217 - Richmond (action level): 209 - Martinez (action level): 204 - Alameda: 194 - Redwood City (action level): 176 Gauges marked action level have no NOAA minor flood level, so they count days over the action level, a watch point. Crossing it doesn’t mean flooding. Counts move as the window rolls forward. Alameda’s was 198 on November 6. Source: Fish Maps station pages, https://tides.andrewlengly.com/stations The week Fish Maps went live, five Bay gauges ran about 1.05 to 1.13 ft above prediction, with a 1.6 ft peak at Alameda on November 3. Predictions use a 1983 to 2001 reference, so part of every gap is sea level rise that has already happened. #### What the data got wrong. NOAA’s own services gave a few wrong answers. - The identify service returned the same polygon at every point - Risk: Every spot reads the same step - Call: Read pixels straight from the level 16 tiles - NOAA’s tiles paint the whole Bay one navy - Risk: The Bay turns navy at every step - Call: Repaint tiles in the browser so open water clears - The terrain tiles include Bay bathymetry - Risk: At 3x the Golden Gate channel becomes a 300 m trench - Call: Set everything below sea level to zero - Only four of seven gauges have a minor flood level - Risk: Flood-day counts with no threshold - Call: Count the action level and label it - First-pass spot points sat on high ground - Risk: China Camp read 8 ft and Point Isabel stayed dry through 10 ft - Call: Move each point to its lowest public road or path - 8.5 − 6.6 comes out to 1.9000000000000004 - Risk: A tide right at the threshold counts as over it - Call: Round to a thousandth of a foot before comparing - USGS imagery bands over open water - Risk: Stripes and gaps across the Bay - Call: Cover the Bay with a flat water tint #### Decisions. Small choices that keep the map honest. - Official dates kept apart: A row gets an official tag only once a Coastal Commission date is entered by hand. The computed peaks never stand in for those dates. - Point Reyes off the Bay calendar: It sits on the open coast and took 13 of the top 20 rows. - A light map: The project brief asked for a light map at #eef3f5, so Fish Maps breaks from the dark maps in the series. - Staffs drawn to be read: A 2 m tide range across a 30 km view is under a pixel at true scale. Staffs scale to the screen, and each label prints the true value. - The map waits on phones: It boots on the first touch or 3.5 s after load. - MapLibre pinned to 5.24: Grid Watch broke on version 6, so Fish Maps follows its pin. #### Design. The water darkens with every foot. Flooded land takes a blue that darkens with each foot of rise, over a light gray-blue map. Coral marks a staff past its flood level or a building in the water. Every step prints its foot value, so color never carries meaning alone. - Type: Schibsted Grotesk for text and IBM Plex Mono for heights and times. - Eleven steps: Lightness falls at every step from 0 to 10 ft, in normal vision and simulated deuteranopia. Color distance between map colors (ΔE), normal vision and simulated deuteranopia (Normal, Deuteranopia): - Water vs flood coral: 93.9, 89.1 - Bay tint vs the 6 ft step: 50.7, 53.6 Neighboring water steps sit 6.6 to 8.2 apart under deuteranopia. The slider prints each foot, and staffs print their height. #### Where it stands. Live, with the king tide season ahead. - 7 Tide gauges - 15 Shoreline spots - 53 Unit tests passing - 0 API keys Lighthouse on mobile, production build (Performance, Accessibility, Best practices, SEO): - Map: 100, 100, 100, 100 - Crab Cove spot page: 100, 100, 100, 100 - Alameda station page: 94, 100, 100, 100 - King tide calendar: 94, 100, 100, 100 The map page’s 100 leaves out the map itself. On phones it boots after Lighthouse stops measuring. With the map in the window, an early build scored 60 on Lighthouse’s software renderer. #### What’s next. What I’d add after the first winter. Next: - Storm scenarios: USGS CoSMoS models storm flooding on top of sea level rise. I found no public map service for it yet. - A phone test: Interaction timings so far come from a laptop GPU with the CPU throttled. - Official dates: Enter the Coastal Commission’s dates by hand. The dates file is still empty, so no row carries an official tag yet. - Checks in CI: Run the tests and the source check on every pull request. So far they run locally. Known limits: - Screening layers: NOAA’s layers leave out storms and waves. A flooded step is a reason to go look. - No weather in predictions: Tide predictions cover astronomy alone. The surge readout shows how far real water is running from them. - No dates on the steps: Sea level rise here is a set of what-if steps. Fish Maps makes no forecast of when any step arrives. #### TL;DR - Built: A 3D map of the Bay that ranks king tides at 7 gauges and draws sea level rise by the foot. - Found: At Alameda, 2 ft of rise tops the minor flood level on 198 predicted days, more than half the year. - By the numbers: 22 Days over Alameda’s minor flood level with 1 ft of rise | 198 Days over it with 2 ft of rise | 341 Days over it with 3 ft of rise - The catch: The rise steps are what-ifs. Fish Maps makes no forecast of when 2 ft arrives. - Next: Add USGS CoSMoS storm flooding on top of sea level rise once I find a public map service for it. ### NetTracer case study Page: https://www.andrewlengly.com/nettracer-casestudy Open NetTracer: https://nettracer.andrewlengly.com Code: https://github.com/AndrewLengLy/nettracer Built with: React, Three.js, Tailwind CSS, Vite, lucide-react NetTracer is a network lab in the browser. It follows one packet across a 3D network a frame at a time and shows which header fields change at each hop. It is built with React, Three.js, Tailwind CSS and Vite, and its 7 guided labs and sandbox are live with the code public. - The idea: Follow one packet across a network and see how the network works. Every frame is on screen, along with the header fields that change at each hop. - What I did: Built an engine that turns a network and a traffic flow into a packet capture, then set it under a 3D scene with 7 guided labs and a sandbox. - Where it stands: Live, with the code public. It has no tests yet, and the header checksum it shows is a stand-in. #### What it does. Seven guided labs and a sandbox run on one engine. - Guided labs: Seven labs, each with a storyline of steps. Three of them start broken, and the capture shows where the packet dies. - 3D stage: Laptops, switches, a router and rack servers on a floor grid. Packets travel the cables as capsules. - Capture table: Every frame in order, with its time, source, destination, protocol and size. Click a row to jump to it. - Step narrative: Each step says why the packet exists and what the device at this hop did with it. - Byte diff: The headers at each layer. A field that changed at this hop shows its old value struck through beside the new one. - Explode OSI Layers: Splits a packet on the stage into its L2, L3, L4 and L7 slabs. - One-click fixes: VLANs, gateway repair, HSRP, TLS and route restore. Each fix re-runs the simulation, so the capture shows what it changed. - Sandbox: Add laptops and servers, then click any device to edit its IP, mask, gateway, VLAN, uplink or power. A traffic generator runs a new capture from the host you pick. #### The questions. Each lab answers one question. - How does a frame reach a MAC address nobody has learned yet? - Lab: 01 ARP & Switch Flooding - In the capture: SW1 floods the ARP broadcast and learns each sender’s port. After that the ping goes out one port - What changes when a packet crosses a router? - Lab: 02 IP Routing & MAC Rewrite - In the capture: R1 writes a new Ethernet frame and the IP addresses never change. Four fields differ after that hop, with TTL going from 64 to 63 - What has to happen before a web page loads? - Lab: 03 HTTP + DNS + TCP Full Flow - In the capture: The DNS lookup and the TCP handshake come first. The GET and its 200 OK follow, 40 frames in all - Why can’t PC-A open any website? - Lab: 04 Broken Gateway Diagnosis - In the capture: Its gateway is 192.168.1.99, an address nobody on the LAN owns. Three ARP requests go unanswered and the packet is dropped - What happens when the only router loses power? - Lab: 05 Router Failure & HSRP - In the capture: ARP for the gateway times out and nothing leaves the LAN. The fix adds a standby router behind a shared virtual IP - Where does a packet go when its route is missing? - Lab: 06 Missing Route to 8.8.8.0/24 - In the capture: R1 has no entry for 8.8.8.0/24, so the DNS query is dropped at the router - How do two VLANs talk over one cable? - Lab: 07 Inter-VLAN Routing on a Stick - In the capture: PC-A in VLAN 10 pings PC-C in VLAN 20 through one trunk to R1. The 802.1Q tag changes from 10 to 20 on the way back out #### How it works. The engine is pure functions, so every edit re-runs the capture. How a network on screen becomes a capture: 1. Network and flow. Each device’s IP, mask, gateway and VLAN. One traffic flow, such as a ping 2. Engine. Pure functions. ARP caches and switch MAC tables. VLAN domains and HSRP election. Longest-prefix routing. Returns: A list of frames, each with the fields that changed since the hop before 3. Three.js stage. Devices and cables built in code. Packets as capsules on the floor 4. React interface. Capture table and step narrative. Byte diff and fixes. Keys: Space plays or pauses, and the arrow keys step one frame The whole app is one React component, 2,960 lines long. The engine sits at the top of the file and needs nothing from the other two parts. A switch records the sender’s MAC against the port a frame came in on. It floods a frame whenever it has no single port to send it to, which covers broadcasts and any MAC it hasn’t learned yet. A router picks the route with the longest matching prefix, and every one of those rules reads the device’s current settings. #### The math. The health score starts at 100 and loses points for each problem. Health score: - health score = 100 minus the penalty of every open issue, never below 0 The header shows the score in green from 85 and in amber from 60. Anything lower is red. Points each issue takes off the health score (Penalty): - A gateway that no router owns: 30 - The only router is down: 30 - A router with no route to its own network: 25 - Two hosts on one IP: 20 - Plain HTTP on port 80: 15 - A single router with no standby: 10 - Every host in one broadcast domain: 10 - A host that bypasses the HSRP virtual gateway: 8 - A router down while the standby carries traffic: 5 The base network starts at 65, with 35 points lost across the issues it ships with. Apply every fix the panel offers and it reads 100. Routing rules: - same subnet = two addresses match once each is ANDed with the mask - route chosen = the matching route with the longest prefix - TTL = starts at 64 and drops by 1 at each router - ARP timeout = 3 requests with no answer Lab devices start on /24 masks. #### What it found. The three labs that start broken score lowest before any fix. The engine runs on its own in Node, with no browser and no 3D scene. The figures below come from running it that way against all seven labs. Health score as each lab opens: - 01 ARP & Switch Flooding (8): 65 - 02 IP Routing & MAC Rewrite (16): 65 - 03 HTTP + DNS + TCP Full Flow (40): 65 - 04 Broken Gateway Diagnosis (7): 35 - 05 Router Failure & HSRP (7): 45 - 06 Missing Route to 8.8.8.0/24 (7): 40 - 07 Inter-VLAN Routing on a Stick (16): 75 Frames in each capture in parentheses. Labs 04 to 06 end in a drop, and each one delivers once its first suggested fix is applied. No lab opens in the green, which starts at 85. #### Where it simplifies. The checksum on screen is a stand-in. The Header checksum field shows a 16-bit slice of an FNV-1a hash over the source, destination, TTL and identification. It changes whenever TTL changes, which is the point lab 02 makes. A real IPv4 checksum is a ones’ complement sum over the whole header, so the value here won’t match a real capture. The README says checksums are computed from live device config. That holds for the inputs, though the arithmetic is a hash. What else is simplified: - UDP checksums - On screen: A checksum on DNS and HSRP datagrams - Underneath: The same hash, taken over the ports and the length - Timing - On screen: A time on every frame, such as 0.015 s for lab 01 - Underneath: Fixed delays. Each hop adds 2 ms or 6 ms, and an ARP request nobody answers adds 1 s - TCP and TLS - On screen: A handshake, then data - Underneath: Scripted packets with fixed ports and sizes. The TLS payload is placeholder bytes - A drop at a router - On screen: The capture ends where no route matches - Underneath: A real router would also send ICMP Destination Unreachable back. The step note says so - HSRP failover - On screen: A Hello from the Active router, and a standby that takes over - Underneath: The powered router with the highest priority owns the virtual IP at once. No hold timer runs #### Decisions. The whole lab ships as one component. - A pure engine: It takes a network and a flow and returns a list of frames. Editing a device runs it again, so the capture always matches the settings on screen. - Hardware built in code: Laptops, switches, the router and the servers are Three.js geometry. The app loads no model files. - One file: All three parts live in one component, with the engine at the top. - Pixel ratio capped at 2: The renderer never draws more than 2 device pixels for each CSS pixel. - Cleanup on exit: Leaving the stage cancels its animation frame and removes its listeners. Geometries, materials, textures and the renderer are all freed. - A note on small screens: Under 1,024 px wide, the app says it was built for a desktop screen and offers an Open anyway button. #### Design. Each protocol keeps one color from the stage to the capture. The interface is dark, with Inter for text and JetBrains Mono for labels and the capture table. Color carries the protocol. - Protocol colors: Eight of them. ARP is cyan, DNS is amber, TCP is blue, and a drop is always red. A packet on the stage and its row in the capture share the color. - Device accents: Each device type has its own accent, shown as the dot on its label. - Camera: A 28° camera looks down on the network like a diorama. Drag to orbit, right-drag to pan, scroll to zoom, and click a device to inspect it. - Playback: Four speeds from 0.5x to 4x, with a strip of ticks that marks every frame. #### Where it stands. Live, with the code public and no tests yet. - 7 Guided labs - 8 Devices in the base network - 9 Issues the health score checks for - 0 Automated tests #### What’s next. What I’d change where NetTracer stops short. Next: - Real checksums: Swap the hash for real ones’ complement checksums on IPv4 and UDP. A real IPv4 checksum still changes with TTL, so lab 02 keeps its point. - Tests and CI: Turn the per-lab figures on this page into the engine’s first tests, and run them on every pull request. The engine already runs without a browser. - A phone view: Put a still of a lab above the Open anyway button, or build a layout for small screens. - A check by hand: Walk the capture rows by keyboard, then play a capture with reduced motion turned on. Known limits: - Checksums: The IP and UDP checksums are hash stand-ins. They won’t match a real capture. - No tests: Nothing checks the engine yet, and the repo has no CI. One pull request has merged since the first build on November 5, 2026, and it added no tests. - Desktop only: There is no phone layout. A small screen gets the note and the Open anyway button. - Keyboard and motion: Capture rows now take keyboard focus, and Up, Down, Home and End move between frames. With reduced motion on, a new capture waits for Play and packets move only when you play or step. The pull request notes list both changes as not checked. #### TL;DR - Built: A network lab in the browser where packets cross a 3D network one frame at a time. - Found: The three broken labs open with the lowest health scores, 35 to 45, and each delivers after its first fix. - By the numbers: 7 Guided labs | 9 Issues the health score checks for | 0 Automated tests - The catch: Nothing checks the engine yet. These scores come from running it in Node, with 0 automated tests behind them. - Next: Swap the hash stand-ins for real ones’ complement checksums on IPv4 and UDP. ### S P L I T case study Page: https://www.andrewlengly.com/tab-split-casestudy Open S P L I T: https://split.andrewlengly.com Built with: Next.js, Supabase, Claude API, Frankfurter S P L I T splits a restaurant bill from one scanned receipt. Friends tap what they had on a shared link, and tax and tip split to the exact cent. It is built with Next.js, Supabase and the Claude API, with Frankfurter for exchange rates, and it is live at split.andrewlengly.com. - The idea: Split the tab without the math. One person scans the receipt, and everyone else taps what they had from a link. - What I did: Built a receipt scanner and a shared bill page on top of split math that adds up to the total. Trips settle many receipts across currencies. - Where it stands: Live at split.andrewlengly.com. 18 tests cover what comes in from receipts and requests, and the split math has none yet. #### What it does. One person scans, and everyone else taps what they had. - Scan: Take a photo of the receipt and Claude reads its line items and totals. You can also type the items in. - Paste: Copy the receipt text with Live Text or Google Lens and paste it in. It’s parsed on the device with no network call. - Who had what: Drag a name from the tray onto an item, or tap a name and then tap items. Split everything evenly covers the whole bill in one tap. - One link: Friends pick their name on the bill link and tap what they had. The bill page refreshes every 3 seconds, so everyone sees the same claims. - Tax and tip: Tax is a rate or an amount, and it falls only on taxed items. Tip is a percent or an amount, split by order or evenly. - Paying back: Each friend sees their amount on buttons for Venmo, Cash App, PayPal and Zelle. The Zelle button copies the handle to paste into a bank app. - Trips: A trip collects receipts in different currencies and converts them to one. It then lists the payments that settle everyone up. - Running tab: The home page keeps a running tab across your recent bills, along with saved groups. Both stay on your device. #### How I built it. Two browsers played host and friend, and every total matched the hand-checked math. From the first pull request to the settle-up fix: 1. First build. Scan, paste, claims, pay buttons and trips. Merged November 3, 2026 2. Host and friend. Two browser sessions in one Playwright run. Claims, live updates, pay links and settle-up. Left out: A real Claude scan and live exchange rates 3. Its own host. One rewrite for every path. Old /split links redirect there. Checked: Seven other hosts served the same pages as before 4. Audit fixes. Each scan logged before the model call. Why: Scans sent together all got past the old limit 5. Settle-up copy. Four places promised the fewest payments. Now they say who pays whom The audit came about a week after launch. A new claim now has to name an item and people that are on the bill, and new bills and trips are capped per visitor. On the bill screen only one save is in flight at a time, so two saves can’t land out of order. #### The questions. Every screen answers a question about who owes what. - Who had what? - Answered by: Claims, one row for each person and item - Shows on: The bill page - What does each person owe once tax and tip are in? - Answered by: Each item shared among the people who claimed it, with tax, fees, discounts and tip spread on top - Shows on: The bill page - Do the shares add up to the bill? - Answered by: Largest remainder rounding to whole cents - Shows on: Every total - Who pays whom after a whole trip? - Answered by: Net balances in one currency, then a greedy settle-up - Shows on: The trip page #### How it works. Each browser does the math from one bill kept on the server. How a receipt becomes a split: 1. Receipt. A photo, shrunk to 2,000 px on its long side. Or text pasted from Live Text 2. Scan route. Posts to the Anthropic Messages API. A Claude Haiku model by default. Reply capped at 2,500 tokens. Every field of the reply checked. Limits: 3 scans per visitor every 30 minutes, and 200 a day across everyone 3. Supabase. Five tables behind row-level security. Reached only from server routes 4. Bill page. Shares computed in the browser. Refreshed every 3 seconds. Also reads: ECB daily rates through Frankfurter, cached 12 hours Pasted text goes to a parser on the device first. It handles rows printed on one line, and the layout where Live Text reads every name before every price. When the result looks off, Have Claude sort it sends the text to the model. The model’s reply never reaches the browser as is. It has to parse as JSON, and each field is read on its own. At most 200 items are kept. The code lives inside this site’s repo, which is private. #### The math. Shares round to whole cents and still add up to the bill. Each person’s share: - item share = item price / people who claimed it - tax share = tax × that person’s taxed items / all taxed items - tip by order = tip × that person’s items / all items - total = items + tax + fees − discounts + tip Fees and discounts follow each person’s part of the item total. A tip split evenly is the tip divided by the number of people. Rounding: - tax from a rate = rate × taxed items, rounded to the cent the way a register does - final shares = floored to whole cents, then the leftover cents go one at a time to the largest remainders A $10.00 pizza split three ways with 7.25% tax and a 20% tip comes to $12.73. One person pays $4.25 and the other two pay $4.24. Trips: - balance = what a person is owed across the trip’s bills minus what they owe, in the trip’s home currency - settle-up = the biggest debt pays the biggest credit, repeated until every balance is zero Only unpaid shares count, and balances settle in whole cents. New US bills start at California’s statewide 7.25%. The picker holds 24 more city and county rates from the CDTFA table effective October 1, 2026, and a tax amount printed on a scanned receipt replaces the rate. Source: CDTFA sales and use tax rates, https://cdtfa.ca.gov/taxes-and-fees/sales-use-tax-rates.htm #### What it found. The shares add up to the cent, and a trip can take one payment more than it needs. The two charts below come from running S P L I T’s own split code, the same functions the bill and trip pages call. Each share of the sample sushi bill: - Zankar: $32.51 - Justin: $31.94 - Nancy: $26.87 - Andrew: $24.66 - Bailey: $17.68 The app’s own sample bill, with Sacramento’s 8.75% tax and an 18% tip split by order. The five shares add up to the bill’s $133.66. Rounding to whole cents left 1 cent over, and it went to Andrew, whose unrounded share of $24.65345 had the largest fraction of a cent. Payments to settle an example trip: - Greedy, as the trip page lists them: 4 - Shortest possible: 3 Two bills shared by five people. A paid the first, and C and D owe $30 each on it. E owes B $50 for the second. Greedy sends E’s $50 to A, so C has to pay both A and B. Paying back whoever covered each bill takes 3. #### What receipts got wrong. Receipt text shows up out of order. - Live Text reads names and prices as separate columns - Risk: Items paired with the wrong prices - Call: The parser handles that layout, and the scan prompt tells the model to pair them up - Pasted text could close its own tag - Risk: Whatever follows reads as instructions to the model - Call: Strip any copy of the tag before sending, and check every reply as data - A long receipt can run past the reply cap - Risk: A reply cut off partway through the list - Call: Spot the cutoff and ask for the items by hand - A burst of scans can outrun the counter - Risk: Requests slip past the limit before they’re counted - Call: Log each scan before the model is called, and stop when the log can’t be written - Some currencies have no ECB daily rate - Risk: A trip total that mixes currencies - Call: Count those receipts at face value and say so on the trip page - Venmo and Cash App only take dollars - Risk: A pay button that opens with the wrong amount - Call: Hide both buttons on any other currency #### Decisions. Nobody signs in, and a bill deletes itself. - No accounts: Starting a bill stores an edit key in that browser, and only that browser can edit the bill. Anyone with the link can claim items and mark themselves paid. - Bills expire: A bill goes 30 days after its last change, or 7 days after everyone has paid the payer back. An empty trip goes after 30 days. - Cleanup with no cron job: Expired bills are cleared when someone starts a new one, at most once an hour. Opening an expired link deletes that bill too. - Polling: The bill page asks the server for the bill every 3 seconds, and only while its tab is visible. Nothing holds a connection open. - No raw IPs: The limit log stores a SHA-256 hash of the visitor’s IP with the kind of request. Rows are deleted after a day. - Caps on what a link can write: A bill holds up to 200 items and 40 people. One visitor can start 30 bills and 10 trips an hour. - Kept out of search: Bill and trip pages are noindex and blocked in robots.txt. #### Design. A name drags onto an item, and a tap does the same job. - Drag pins: Names drag from the tray with a mouse or a finger. Drag a pinned name to another item to move it, or back to the tray to take it off. - A short press on phones: A name already pinned to an item needs a 260 ms press before it lifts, so a swipe down the list still scrolls. - Scrolling mid-drag: Hold a name near the top or bottom edge and the page scrolls under it. - A tap path: Tap a name, then tap items. It does what the drag does. - A color for each person: Names take one of 12 colors, in the tray and on every item they’re pinned to. - Scoped styles: Every rule sits under .tabsplit, so none of it reaches the rest of the site. #### Where it stands. Live, with tests on what comes in and none on the math. - 18 Tests passing - 0 Tests on the split math - 27 Currencies a bill can use - 25 California tax rates built in What the tests cover (Tests): - Receipt text parser: 4 - The model’s scan reply: 8 - Claim requests: 6 - Shares and rounding: 0 - Trip settle-up: 0 The tested code is the code that takes outside input. S P L I T was added to this site on November 3, 2026 and now runs on its own subdomain. #### What’s next. Tests on the money math come first, then an exact settle-up. Next: - Math tests: Cover the share math and the settle-up, starting with the examples on this page. - Shortest settle-up: Replace the greedy method with one that finds the shortest list of payments. The example trip above is the first test it has to pass. - A real scan: Run a real receipt through Claude and a trip on live ECB rates. The first test had neither. - Quarterly rates: Check the built-in California rates against the CDTFA table when each new quarter starts. Known limits: - Untested math: The share math and the trip settle-up have no tests. The examples on this page each come from a single run of the code. - Greedy settle-up: The list of payments is not always the shortest one. The method is greedy, and the example trip above takes 4 payments when 3 would do. - Tax rates by hand: CDTFA changes rates each quarter. The built-in list is dated October 1, 2026 and gets updated by hand. - One browser holds the key: The edit key lives in one browser’s storage, and there is no sign-in to get it back. - Face value: A receipt in a currency with no ECB rate counts at its printed number on a trip. #### TL;DR - Built: A receipt scanner and bill link where friends tap what they had and tax and tip split to the cent. - Found: The sample bill’s five shares add up to its $133.66, and the example trip takes 4 payments where 3 would do. - By the numbers: $133.66 Sample bill, split to the cent | 18 Tests passing | 0 Tests on the split math - The catch: The split math has no tests. Each example on this page comes from a single run of the code. - Next: Write tests for the share math and the settle-up, starting with the examples on this page. ### Campus IT Triage Agent case study Page: https://www.andrewlengly.com/campus-it-casestudy Open Campus IT Triage Agent: https://github.com/AndrewLengLy/Campus-IT-Triage-Agent Read the architecture notes: https://github.com/AndrewLengLy/Campus-IT-Triage-Agent/blob/main/docs/architecture.md Release v1.0.0: https://github.com/AndrewLengLy/Campus-IT-Triage-Agent/releases/tag/v1.0.0 (Explainer video) Built with: Salesforce DX, Apex, Service Cloud, Agentforce, Custom Metadata Campus IT Triage Agent is a Salesforce Service Cloud backend for a help desk agent that tries a known-issue guide first and opens a Case when it is needed. A chatbot that only searches a FAQ opens duplicate tickets. This one is built as five Apex actions for an Agentforce agent, and it is public on GitHub with 29 test methods. - The problem: Campus IT desks get the same requests every shift. A chatbot that only searches a FAQ opens duplicate tickets and can’t show what it settled. - What I did: Built five Apex actions an Agentforce agent can call, with routing and response times kept in Custom Metadata. Every action writes a row to an audit object that a trigger keeps append-only. - Where it stands: Public on GitHub with 29 test methods. The chat layer isn’t in the repo, and the walkthrough video is a reconstruction. #### The problem. A campus IT desk answers the same requests every shift. The repo’s README opens with the problem. Campus service desks get WiFi, password, MFA, VPN and Outlook requests every shift, and hardware that needs a technician waits behind work a script could finish. A chatbot that only searches a FAQ makes that worse. It opens duplicate tickets, and nobody can prove whether it deflected anything. The backend here is built around those two failures. Source: The README, https://github.com/AndrewLengLy/Campus-IT-Triage-Agent#the-problem Who runs it: - Service desk lead: Owns the known-issue articles and the rules for queues and response hours. - Technician: Works Cases in the Campus IT app. - Agentforce admin: Owns the topic instructions and which actions the agent can call. - Platform owner: Handles deploys and tests, and assigns the permission set. #### What it does. Five actions, and each one does a single job. The five actions: - Find a guide: Matches the request against six known-issue articles. A match returns the steps, and no Case is opened. - Escalate a ticket: Opens one Case on the queue for its category and links the student’s Contact. The Case carries a published comment and a first-response due time. - Add an update: Adds a comment to an open ticket. A closed ticket is refused, and the agent is told to open a new one. - Check status: Looks a ticket up by Case Number, and falls back to Student ID when the number is wrong. - Operations snapshot: Reports the open agent-sourced tickets, how many are high priority, the oldest wait, a count for each queue and today’s deflection rate. Two behaviors built in: - One ticket per problem: An open Case for the same student and category is reused. Force New Ticket overrides that. - Lookups that survive speech: A Student ID given as s 1000-0001 is stored as S10000001, and Case Number 1234 matches 00001234. #### How I built it. I built the backend first, then checked the walkthrough against the code. All six commits on main, in five steps: 1. First build. The backend and all 29 test methods in one commit. Tagged: Release v1.0.0, with an 80-second explainer video 2. Public repo. A packaging guide and launch copy. A script for recording after a deploy 3. Demo video. A page rebuilt to look like the workspace. No network calls, and every agent reply scripted 4. Honest labels. The opening caption had called it a live desk. Added: “UI walkthrough · not a live org capture” on every caption 5. Audit, written down. Nine steps, starting with claims the code doesn’t back. A voice check the launch post passed The label stays on screen for all 65 seconds of the walkthrough, and the README says the same. Source: The pull request that fixed the labels, https://github.com/AndrewLengLy/Campus-IT-Triage-Agent/pull/2 #### The questions. The snapshot answers what a desk lead would ask. - How many agent-sourced tickets are open? - Measure: Open Cases with Agent Sourced checked - Read from: Case - How many of those are high priority? - Measure: Open agent-sourced Cases at High priority - Read from: Case - How long has the oldest one waited? - Measure: Whole hours since the oldest open ticket was created - Read from: Case - Which queue is holding them? - Measure: Open tickets counted by the queue that owns them - Read from: Case and its queue - What share of today’s requests ended in chat? - Measure: Deflection rate - Read from: Campus IT Interaction rows created today #### How it works. This repo is everything the agent can do, minus the conversation. The path a student request takes: 1. Find a guide. Keywords matched against the catalog. Steps read back to the student 2. Escalate. Reuse an open Case or insert one. Stamp Agent Sourced and the due time. Publish a Case Comment. Route to a Campus IT queue. Only when: The guide is missing, already tried, marked for escalation, or the student asks for a person 3. Update and status. Comment on an open ticket. Look up by Case Number or Student ID 4. Snapshot. Open load by queue. Today’s deflection rate. Every action: Writes one Campus IT Interaction row Each action is one Apex class with a single invocable method that takes a list and returns a list. All five run with sharing. The project deploys as Salesforce DX source on API version 64.0. The conversation is built in Agent Builder inside an org, so it isn’t in the repo. #### The math. Deflection is the share of today’s requests a guide settled. Deflection today: - deflection rate = guides matched today / guides matched today + new tickets created today A reused ticket counts as a follow-up and stays out of the bottom line. The rate is a percent to one decimal place, and it stays empty on a day with no volume. Open load: - open load = open Cases with Agent Sourced checked, for the whole desk or one queue - oldest wait = whole hours since the oldest of those Cases was created First response due: - High = 4 hours after the Case is created - Medium = 16 hours - Low = 40 hours - no matching rule = 16 hours These are clock hours added to the moment the Case is created. Nights and weekends don’t pause them, and changing a rule leaves open tickets as they were. Priority from what the student says: - High = the urgency mentions critical, emergency, urgent, high, p1, p2 or sev1 - Low = the urgency mentions low, or the codes p4 and sev4 Everything else is Medium, including a request with no urgency given. #### What it found. One test checks the deflection rate, and two of its rules go untested. Test methods by class on main, 29 in all: - Ticket escalation: 15 - Ticket status: 4 - Audit service: 3 - Ticket update: 3 - Self-service guide: 2 - Demo data: 1 - Operations snapshot: 1 Counted from the code, since no test run is on record. The snapshot test checks the rate against one mix of matched guides and new tickets. No test puts a reused ticket in that mix or runs the snapshot on a quiet day. Source: The snapshot test, https://github.com/AndrewLengLy/Campus-IT-Triage-Agent/blob/main/force-app/main/default/classes/ITDeskOperationsSnapshotTest.cls #### What’s staged. The walkthrough video is a reconstruction, and it says so on screen. The repo carries a 65-second walkthrough of the Campus IT workspace. It mirrors what the Apex does and was not captured from a live org. The README says so, and the video is labelled on screen. Any count that appears in the video belongs to the reconstruction. Every figure in this study comes from the code and the metadata. In the repo: - Apex: 15 classes. Five actions, two shared services, a demo data seeder and seven test classes. - Metadata: Four custom objects and metadata types, holding 14 Custom Metadata records. Five queues take the routed Cases. - Workspace: A Lightning app, a Case layout, list views and one permission set. - Demo data: A script that seeds three demo students, one for each walkthrough path. Not in the repo: - The conversation: The Agentforce topic is built in Agent Builder inside an org. The repo carries the instructions to paste in. - A live recording: No capture from a real org exists here. A script for recording one is in the docs. - An install link: The README keeps a placeholder for a package URL and says not to invent one. - A test run: No test result or coverage figure is recorded. #### Decisions. The rules live where an admin can change them. - Agent copy stays in Agent Builder: A desk lead can change the wording with no deploy. - Rules in Custom Metadata: Queue routing and response hours change in Setup with no Apex. So does the article catalog. - One action, one job: Each class takes a list and returns a list in the same order, so it stays safe in bulk. - The audit never blocks a student: Audit writes run in system mode and catch their own errors. A logging failure can’t fail the action a student is waiting on. - Append-only by trigger: A trigger rejects every update and delete on the interaction object, so the numbers can’t be edited afterward. - Tests bring their own data: No test uses SeeAllData. Test queues are named CIT_Test_* so they can’t collide with the shipped ones. - No hardcoded IDs: Queues are found by DeveloperName, and no Salesforce record ID appears in the code or the metadata. - No secrets: The project needs no API keys and no OAuth client. Deploying takes the Salesforce CLI and one permission set. #### Design. Technicians get a workspace that works without the agent. - Campus IT app: A Lightning app for Cases, Contacts, Interactions and their due dates. - Case highlights: An optional Case layout puts Student ID and first-response due in the highlights. - List views: Four of them. Two cover Cases and two cover the interaction log, with one scoped to today for a shift review. - Interaction numbers: Each audit row carries an autonumber in the form INT-00000. #### Where it stands. The code is public, and no run in a real org is on record. - 5 Agent actions - 29 Test methods in 7 test classes - 14 Custom Metadata records - 6 Known-issue articles The repo is MIT licensed, with release v1.0.0 and six commits on main. The first is from August 2026. Two pull requests are open, and one of them adds more tests and a one-command scratch org setup. #### What’s next. First, it needs a run in a real org. Next: - Tests in an org: Run the seven test classes in a scratch org and publish the coverage figure. An open pull request adds a script that sets one up and runs every test with coverage in one command. - More tests: The same pull request adds tests that push 200 records through every action in one transaction, and one for the empty rate on a quiet day. A test with a reused ticket in the rate still needs writing. - Open decisions: Writing those tests turned up 13 issues, each listed in the pull request for a decision. In one, nothing decides which queue wins when two routing rules match. - A live capture: Record the walkthrough from a Developer Edition org with Agentforce enabled. Then the on-screen label can come off. - Business hours: Work out first-response due times with BusinessHours.add() against the desk’s Business Hours record, so nights and weekends pause the clock. - A classifier: Keep keyword matching as the floor and add a classifier above it for requests worded another way. Known limits: - No run on record: The repo records no test run or coverage figure. Nothing in it shows a deploy to a real org. - A reconstructed video: The walkthrough is a labelled reconstruction. The script for recording a real one is in the docs. - Agentforce needs the right org: A standard Developer scratch org doesn’t have Agentforce. The chat demo needs a Developer Edition org or a Trailhead Playground. - Clock hours: First-response times ignore business hours. - Keyword matching: Guides match on keywords from the catalog. A request worded another way finds no guide and goes to escalation. #### TL;DR - Built: A Salesforce Service Cloud backend whose campus IT agent tries a known-issue guide before it opens a Case. - Found: Of the 29 test methods, one checks the deflection rate, and two of the rate’s rules go untested. - By the numbers: 5 Agent actions | 29 Test methods in 7 test classes | 6 Known-issue articles - The catch: No test run is on record. The 29 test methods are counted from the code. - Next: Run the seven test classes in a scratch org and publish the coverage figure. ### AlwaysUp case study Page: https://www.andrewlengly.com/alwaysup-casestudy Open AlwaysUp: https://alwaysup-mu.vercel.app Code: https://github.com/AndrewLengLy/AlwaysUp The flip in code: https://github.com/AndrewLengLy/AlwaysUp/blob/main/src/lib/flip.ts Built with: React, Vite, TypeScript, Framer Motion, Tailwind CSS, Vitest, Claude Code AlwaysUp is a joke portfolio viewer that mirrors a losing chart about the price you paid, so a loss is drawn as a gain of the same size. It is a parody about how a red number can ruin a morning, and it says so on every screen. It is built with React, Vite and Framer Motion, and it is public on GitHub with 77 unit tests. - The idea: Checking a portfolio costs me more in mood than it does in money. So I built an app that won’t show me a loss. - What I did: Wrote the flip as one line of math, made it something you can tap, drag and hold, and labeled every screen with what was done to it. - Where it stands: Live and public on GitHub. It’s a joke, and nobody should invest with it. The live deploy runs on simulated prices. #### Why I built it. A red number can ruin a whole morning. The hardest part of investing, for me, isn’t picking anything. It’s what a red number does to the next hour, over a position I wasn’t going to touch anyway. The position is the same whether I look or not. I look anyway. So I made an app that shows the number going up, every time, and is completely upfront that it’s lying. #### What it does. Losing charts get flipped so the line climbs. The whole trick: - p′ = 2a − p, a reflection of every price p about an anchor a The anchor is the price you paid, or the opening price when you’re still in profit. The shape stays exactly the same and the direction inverts, so a 6.37% slide becomes a 6.37% climb. The four cases: - Up today, up overall - Does: Nothing. Nothing on screen is a loss. - Down today, up overall - Does: Mirrors the day about the open. - Down today, down overall - Does: Mirrors about what you paid. One flip fixes the day and the position together. - Up today, down overall - Does: Lifts the line clear of what you paid. A mirror would throw away the green day. Three comfort levels: - Honest: No flip. Red is red. - Comfort: Losses are mirrored and gains are left alone. This is the default. - Delulu: Every chart climbs. Down moves are folded upward and damped, so the line never loses ground. #### Keeping it a joke. It lies on purpose, so it has to say so everywhere. AlwaysUp is a parody. It isn’t investment advice, it can’t trade, and it never tells anyone what to buy. A chart that looks real and isn’t could still fool someone, so the app keeps reminding you what it is. How it keeps reality one tap away: - A banner on every screen: “A joke app. The charts are flipped on purpose. Don’t invest with it.” - A disclosure first: Before any number shows, you have to accept a screen that opens with “This app lies to you on purpose.” - A badge that names the trick: The header says Mirrored, Lifted or Showing reality, and it can’t be dismissed. - Flip it back anytime: Tap a chart, drag the mirror track or press and hold to see what really happened. - Screen readers get the truth: Every chart carries the real price action and the real profit or loss for assistive tech. - Fake data is labeled: When real prices can’t be fetched, every invented one carries a Simulated tag in every mode. #### How I built it. I built the flip first, then made it something you can play with. The eight commits on main, in five steps: 1. First build. The mirror on fictional tickers. Simulated prices on a seeded walk 2. Real companies. Live prices from Yahoo Finance. Tap, drag and hold to flip 3. A badge. The parody stickers came off. Added: A badge that names the operation in force 4. What you paid. The mirror anchored on your cost. A lift for the case a mirror can’t fix 5. Demo video. A 61-second reconstruction. Simulated prices, labeled as such The restyle to match this site came last, along with the banner on every screen, in pull request #2. Source: Pull request #2, https://github.com/AndrewLengLy/AlwaysUp/pull/2 #### What it found. Flipping was easy. Knowing when a flip is wrong took the work. A reflection inverts every direction. On a green day for a position that’s still underwater, a mirror would turn the good day red to fix the total. That case gets a lift, and the badge says Lifted, not Mirrored. Unit tests by file on main, 77 in all: - The flip: 38 - Chart geometry: 14 - Yahoo parsing: 11 - Position math: 8 - Number formats: 6 Source: The flip tests, https://github.com/AndrewLengLy/AlwaysUp/blob/main/src/lib/flip.test.ts #### What’s next. Real prices on the live site, and the joke stays loud. Next: - Real prices live: Yahoo Finance only works through the dev server’s proxy. The README names Alpha Vantage as the source that works on a static deploy. - Keep the labels: Any new screen gets the banner and the badge before it ships. Known limits: - Simulated live: The live deploy has no proxy, so every price there is simulated and tagged. - Made-up positions: The starting share counts and costs are invented. Settings says so, and you can enter your own. - Delulu isn’t real: Delulu rebuilds the series, so what it draws never happened. - Saved in one browser: Holdings and the comfort level live in local storage. #### TL;DR - Built: A joke portfolio viewer that mirrors a losing chart about what you paid, so the line always climbs. - Found: A mirror can’t rescue one of the four cases, a green day on a position still underwater, so that one gets lifted. - By the numbers: 1 Line of math behind the flip | 4 Cases the anchor has to handle | 77 Unit tests - The catch: The live deploy can’t reach Yahoo Finance, so every price there is simulated, and each one is labeled. - Next: Give the live deploy a data source that works without a proxy, so its prices are real. ### The case for third places case study Page: https://www.andrewlengly.com/third-places-casestudy Open the full study: https://analyst.andrewlengly.com/studies/third-places Built with: Python, pandas, statsmodels, SciPy, geopandas, networkx, esda, Plotly A series of seven data studies on loneliness and the places people gather, run to decide where FourthSpace launches in Sacramento. It runs survey-weighted models on federal microdata, including 540 versions of the social media link, and maps walk access to indoor venues across Sacramento County. Built with Python, pandas, statsmodels and geopandas, it recommends a launch in south Sacramento with library branches as the first venues. - The question: Do people need more places to meet in person, and where in Sacramento should FourthSpace start? - What I did: Ran seven studies on time diaries, CDC surveys, county business counts and a walk network I built for Sacramento. Every figure went into one stat bank with its source. - Where it stands: Done in November 2026. I picked south Sacramento for the launch, with library branches as the first venues. The findings also went into the FourthSpace pitch deck. #### The question. An app for meeting people in person only works if people want to meet. FourthSpace is a community app I’m building for finding third places that are open right now. Think coffee shops, libraries, trails and farmers markets. It launches in Sacramento. Before launch I wanted the data to settle where in Sacramento it starts and who it starts with. I came in thinking phones and social media push people into isolation. I was partly right. #### How I did it. Seven studies on public microdata, with every number kept in one stat bank. From ten questions to a launch pick: 1. Scope. 10 questions scoped. 7 run, each tied to a launch call 2. Microdata. ATUS diaries, 2003 to 2025. CDC YRBS and PLACES 3. Weights. Each survey’s design weights. Taylor-linearized variance 4. Every model. 540 social media models. The full spread kept. Why: One model can be tuned toward an answer, so I ran every reasonable one 5. On foot. Walk network from TIGER roads. Freeways and ramps removed 6. Stat bank. Year, source and caveat per figure. A flag for landing page use - walk access = the share of a tract’s street intersections within 800 meters of an indoor gathering place, about a 10-minute walk - priority tract = a census tract with higher loneliness and low walk access - lonely adults = the CDC PLACES crude rate times the adults who live there, an estimate built on a model Before writing any finding, I checked my weighted rates against the figures the CDC and Census publish for the same year. What I fixed along the way: - Bookstores at zero: The 2022 and 2023 business files still list bookstores under the 2017 industry code. Reading them with the new code matched nothing, so I kept the 2017 code for every year. - Winding roads counted twice: My first access measure counted every point along a road’s shape. It now counts intersections only. - An institution on top: A special land-use tract topped my first ranking. Tracts numbered 98xx and 99xx are now out of every ranking. - A wrong place name: My first map caption put the hot spots in the northern suburbs. Matched to ZIP codes, they sit in south and north Sacramento. #### What the data shows. Time with friends fell by more than half since 2003. Minutes a day with friends, Americans 15 and up: - 2003: 60.0 - 2008: 50.2 - 2013: 54.8 - 2019: 34.3 - 2020: 19.7 - 2025: 26.5 Weighted American Time Use Survey diary means. Ages 15 to 24 fell from 155 to 54 minutes, and everyone spent 49 more minutes a day alone. 2020 was the pandemic low. Extra points of persistent sadness, U.S. high school students, 2025: - Bullied electronically: +31.3 - Bullied at school: +27.8 - Does not feel close to people at school: +17.3 - Sleeps under 8 hours on school nights: +14.1 - Plays on no sports team: +10.7 - Uses social media more than once an hour: +8.3 Each bar is its own survey-weighted model with demographic controls. Across all 540 social media models, 485 put heavy users at higher risk, with a median gap of 5.0 points. Heavy social media use does go with more sadness. Bullying and weak ties at school go with far larger gaps. Each bar is an association from one survey wave, and the direction is open too, since sad teens may turn to social media more. #### Sacramento. Sacramento County is lonelier than 95% of U.S. counties. - 39.5% Adults lonely at least sometimes, 2023 - 2 of 58 California counties, loneliest first - ≈477K Lonely adults in the county, my estimate - 36 Priority tracts Adults who feel lonely at least sometimes, age-adjusted, 2023: - Sacramento County: 39.5% - California county median: 36.2% - U.S. county median: 34.0% CDC PLACES model-based estimates. The 36 priority tracts hold about 52,000 lonely adults, and 22 of them sit inside a loneliness hot spot. Of those tracts, 34 have no indoor venue mapped at all. Most of the hot spot tracts sit along the Florin Rd and Stockton Blvd corridors in south Sacramento. #### Where people gather. Civic clubs per person fell by 40% in 25 years. U.S. establishments per resident in 2023, 1998 = 100: - Coffee shops and snack bars: 191 - Fitness and recreation centers: 151 - Barber shops: 143 - Full-service restaurants: 110 - All establishments: 99 - Religious organizations: 95 - Bars and taverns: 62 - Civic and social organizations: 60 - Bookstores: 31 - Bowling centers: 28 Census County Business Patterns. In Sacramento County, civic and social organizations per resident fell 34%, and 92 remain with paid staff. Minutes with anyone outside the household on a workday, 2021 to 2025: - Working from home: 57 - At the workplace: 423 U.S. employed people on days with 4 or more hours of work. Most of the gap is coworkers. Time with friends is about the same on both kinds of day. Americans spent 83 minutes a day at third places in 2003 and 67 in 2025. Time at home grew by 74 minutes a day over the same years. #### My recommendation. Launch in south Sacramento, with the library branches as the first venues. - A. Start in the loneliest neighborhoods - Why: 22 of the 36 priority tracts sit inside a loneliness hot spot, along Florin Rd and Stockton Blvd - Risk: 34 of those tracts have no indoor venue mapped, so early users may find few places nearby - B. Make the library the anchor partner - Why: Sacramento Public Library logged 2,010,888 visits in fiscal 2024, 54% of its fiscal 2019 count, and its branches are free and run their own events - Risk: Public agencies take time to approve partners - Together: The Southgate and Valley Hi-North Laguna branches give early users in those tracts somewhere to go If library approval runs long, the branches can still be listed as public places while the partnership is worked out. The research also became a 12-slide pitch deck and a ranked list of venues to call. #### Limits. What this data can’t tell you. - Cause: Every link between loneliness, phones, places and health in these studies is an association. - Modeled rates: PLACES fills in counties and tracts from a statewide model, so the 477,000 and 52,000 figures are estimates built on estimates. - Counts and minutes: Business counts and diary minutes say nothing about how a place feels or whether people meet anyone there. - Map gaps: Walk access depends on OpenStreetMap. A venue missing from the map makes a tract look emptier than it is. #### What’s next. What I’d add with local and first-party data. - Local teens: Ask the California Healthy Kids Survey for Sacramento County results, to see whether the school ties pattern holds here. - Two waitlist questions: Add optional questions on loneliness and volunteering to the FourthSpace waitlist, to build a first-party measure over time. - Foot traffic: Test Advan or SafeGraph visit data in the priority tracts. It costs money, so it stays optional. - Remote work again: Rerun the workday finding with the 2026 ACS once it shows what the return-to-office order did. #### TL;DR - Studied: Where my community app FourthSpace should launch in Sacramento, worked out across seven studies on loneliness. - Found: Americans 15 and up went from 60.0 minutes a day with friends in 2003 to 26.5 in 2025. - By the numbers: 26.5 Minutes a day with friends, 2025 | ≈477K Lonely adults in the county, my estimate | 36 Priority tracts - The catch: The Sacramento loneliness rates come from a statewide model, so the local counts are estimates built on estimates. - Next: Ask the California Healthy Kids Survey for Sacramento County results to test the school ties pattern locally. ### Where the best hires come from case study Page: https://www.andrewlengly.com/sourcing-channels-casestudy Open Where the best hires come from: https://www.andrewlengly.com/studies/sourcing-channels.pdf Built with: Python, pandas, lifelines, SciPy, DuckDB, Jupyter, BLS API A data study of which hiring channels bring in people who perform well and stay two years, tracking 311 hires by source. It runs survival models on retention and compares performance ratings by channel. Built with Python, pandas, lifelines and DuckDB, it recommends a two-tier referral bonus and a separate channel mix for production and office roles. - The question: Referrals, job boards, LinkedIn, job fairs and search ads all bring in hires. Which ones bring in people who stay and do well? - What I did: Tracked 311 hires by source through retention, survival models and performance ratings, then checked the pattern against BLS quit rates and published hiring data. - Where it stands: Done in November 2026. I recommend a two-tier referral bonus and a separate channel mix for production and office roles. #### The question. A recruiting budget is a bet on which channel brings in people who stay. Most teams measure channels by cost per hire. That number stops at the offer letter. A cheap hire who leaves in month eight costs the full search again. So I measured each channel by what happens after the hire. Who is still there at two years, and how do their managers rate them? #### How I did it. One table of hires, followed from the offer to the latest review. From raw HR records to a channel mix: 1. Clean. 311 hires, 2006 to 2018. Tenure to the last review date 2. Retention. Kept at 12 and 24 months. Wilson intervals, Fisher test 3. Survival. Kaplan-Meier curves. Cox model with controls. Controls: Role family and hire year, since earlier hires had longer to leave 4. Cost. Cost per hire still there at 24 months. Editable cost table 5. Benchmark. BLS JOLTS quit rates. Published referral research - kept at 24 months = still employed two years after the hire date, counted only for people hired at least two years before the last review - cost per retained hire = cost per hire / share kept at 24 months Channel costs are assumptions. Referrals assume a $2,500 bonus plus about $1,000 of recruiter time. Every other channel uses SHRM’s published average of $4,700. #### What the data shows. Referral hires stayed longest and were rated below expectations least. Still employed at 24 months, channels with 20+ hires: - Employee referral (31): 97% - LinkedIn (76): 95% - CareerBuilder (23): 91% - Diversity job fair (29): 90% - Indeed (84): 87% - Google search ads (49): 84% Hire counts in parentheses. Huebner HR Dataset, synthetic. Hires rated Needs Improvement or put on a PIP: - Employee referral: 3% - LinkedIn: 8% - Google search ads: 8% - Indeed: 10% - Company website: 15% - Diversity job fair: 17% - CareerBuilder: 17% Cost per hire still there at 24 months, assumed costs: - Employee referral: $3,617 - LinkedIn: $4,961 - CareerBuilder: $5,148 - Diversity job fair: $5,242 - Indeed: $5,408 - Google search ads: $5,617 A referral stays the cheapest until it costs about $4,548 per hire. That leaves room for a bonus of about $3,500. The gaps all point the same way, and the sample is too small for any one of them to pass a significance test. Referral hires kept 97% at two years against 90% for everyone else (Fisher p = 0.33). The Cox model puts their risk of leaving 22% below Indeed hires, with an interval from 0.30 to 1.99. Published data at far larger scale shows the same pattern. On Ashby’s platform referrals were about 1% of applications and about 18% of hires, and hiring managers scored them highest. Source: Referral statistics, with Ashby and New York Fed figures, https://www.noon.ai/blog/articles/360-employee-referral-statistics-2026 #### By role family. Referrals held up everywhere. The runner-up changed with the job. - Production roles - Best: Referrals and CareerBuilder kept 100% at two years, LinkedIn kept 95% - Watch: Indeed kept 83%, and 8% of its production hires quit in year one - Office roles - Best: The careers page kept all 9 of its hires. LinkedIn and referrals each kept 93% - Watch: Google search ads kept 80% from a small base BLS puts manufacturing’s annual quit rate at 23% on average from 2021 to 2025. Every channel here beat that in year one, so this workforce is steadier than the national average. Source: BLS Job Openings and Labor Turnover Survey, https://www.bls.gov/jlt/ #### My recommendation. Pay for referrals that stay, and split the mix by role. - Pay the bonus in two parts: Half at 90 days and half at 12 months, so the bonus rewards hires who stay. - Two tiers: About $1,500 for production roles and up to $3,500 for office and technical roles, under the break-even point. - Production mix: Referrals and LinkedIn first, CareerBuilder as backup, and a smaller share for Indeed. - Office mix: Referrals and LinkedIn first. The careers page earns a test with more volume. - Search ads last: Lowest two-year retention (84%) and the fewest Exceeds ratings (4%). - Track it quarterly: Two-year retention and first-year rating by source, with the mix revisited once a year. #### Limits. What this data can’t tell you. - Synthetic data: The Huebner dataset was built for teaching HR analytics. Its numbers show the method, and the published research carries the conclusion. - One source per hire: Someone referred by a friend who then applied on Indeed may be logged under Indeed. - Assumed costs: A real employer should swap in invoices and recruiter hours. #### What’s next. What I’d run with a real company’s data. - A bonus test: Roll the two-part bonus out to half the departments for two quarters and compare referral volume and 12-month retention. - Source of influence: Ask new hires where they first heard about the job, beside where they applied, to catch referrals logged under job boards. - Quality of hire score: Blend 12-month rating, retention and time to first independent work into one score per channel. - Referrer retention: Check whether employees who refer someone also stay longer, a pattern iCIMS reported. - A live dashboard: Rebuild the cleaned tables in Power BI so recruiting can watch channel quality every month. #### TL;DR - Studied: Which hiring channels bring in people who perform well and stay two years, tracked across 311 hires. - Found: Referral hires stayed longest, with 97% still employed at 24 months against 90% for everyone else. - By the numbers: 97% Referral hires still employed at 24 months | 3% Referral hires rated Needs Improvement or put on a PIP | $3,617 Referral cost per hire still there at 24 months - The catch: The sample is too small for any gap to pass a significance test. The referral gap has p = 0.33. - Next: Roll the two-part referral bonus out to half the departments for two quarters and compare retention. ### Skills-based hiring case study Page: https://www.andrewlengly.com/skills-hiring-casestudy Open Skills-based hiring: https://www.andrewlengly.com/studies/skills-hiring.pdf Built with: Python, pandas, DuckDB, Jupyter, Excel, O*NET A data study of how many jobs ask for a degree their own workers do not hold, across 177 occupations that list a bachelor's degree. It joins BLS data to O*NET's worker surveys and builds a launch list of ten roles. Built with Python, pandas, DuckDB and Excel, it recommends skills-first job templates shipped together with built-in assessments. - The question: Employers keep announcing they’ve dropped degree requirements. Why has so little changed in who they hire, and what can an ATS do about it? - What I did: Joined BLS data on 177 bachelor’s-listed occupations to O*NET’s worker surveys, sized the openings, and built a launch list of ten roles. - Where it stands: Done in November 2026. I recommend skills-first job templates shipped together with built-in assessments. #### The question. Companies dropped the degree line. Most kept hiring the same people. The Burning Glass Institute and Harvard Business School followed 11,300 roles at large firms before and after degree requirements came off the posting. Non-degree hiring rose 3.5 points on average. That touched fewer than 1 in 700 hires in 2023. Greenhouse sits where that decision gets made, in the job post and the screen. So I asked where the degree requirement is weakest and what product would make the change stick. Source: Skills-Based Hiring: The Long Road from Pronouncements to Practice, https://www.burningglassinstitute.org/research/skills-based-hiring-2024 #### How I did it. Two government sources, joined on the occupation code. From BLS and O*NET tables to a launch list: 1. BLS tables. Typical entry education. What current workers hold. Jobs and openings 2. O*NET. What workers say the job requires. Skill importance 3. Join. 6-digit SOC code. 742 occupations matched. Detail: O*NET splits some codes into .01 and .02, so values average up to the 6-digit code 4. Size. Openings times the share without a degree 5. Shortlist. Four filters. Ranked by openings - degree gap = the share of workers 25 and older in a bachelor’s-listed job who hold no bachelor’s (BLS Table 5.3) - skills-first openings = annual openings times the degree gap - distinctive skill = a skill whose importance in this job sits furthest above its average across all jobs (O*NET) Generic skills like active listening rank high in nearly every job, so the shortlist shows distinctive ones. Those make better assessment targets. #### What the data shows. In bachelor’s-listed jobs, 3 in 10 workers have no bachelor’s. - 177 Occupations listing a bachelor’s - 12.7M Workers in them without one - 970K Skills-first openings a year - 0.48 BLS and O*NET agreement No bachelor’s, in bachelor’s-listed jobs with 100,000+ workers: - Construction managers: 65% - Cost estimators: 61% - Facilities managers: 61% - Buyers and purchasing agents: 57% - Logisticians: 55% - General and operations managers: 53% - Technical sales representatives: 50% BLS Table 5.3, 2023 to 2024. Postsecondary teaching assistants (66%) are left off, since schools set their own rules. O*NET asks workers what education their job needs. Across bachelor’s-listed occupations, the share who say less than a bachelor’s is enough tracks the share who hold less than one (correlation 0.48). 30 occupations clear 30% on both. #### Why templates need tests. The firms that changed who they hired changed how they judged skill. What firms that dropped degree requirements did next: - Changed who they hired: 37% - Changed the posting only: 45% - Gained, then slid back: 18% Source: Burning Glass Institute and HBS, 2024, https://www.burningglassinstitute.org/research/skills-based-hiring-2024 At the firms that did change, non-degreed hires stayed at a rate 10 points higher than their degree-holding colleagues. A 2025 follow-up across 1,000+ employers found that removing the requirement alone added about 2 points of non-degree hires. The leaders changed how candidates got screened. An assessment built into the ATS is that change. #### Launch list. Ten roles to launch with, about 260,000 openings a year. Launch roles ranked by skills-first openings (Openings a year, No bachelor’s (BLS), Say less is needed (O*NET)): - General and operations managers: 285.0K, 53%, 72% - Construction managers: 49.7K, 65%, 55% - Technical sales representatives: 23.6K, 50%, 63% - Computer occupations, all other: 27.0K, 43%, 42% - Compliance officers: 32.7K, 35%, 37% - Administrative services managers: 22.0K, 43%, 72% - Facilities managers: 14.2K, 61%, 54% - Industrial production managers: 17.0K, 50%, 53% - Loan officers: 17.1K, 46%, 30% - Network and computer systems administrators: 13.4K, 47%, 38% Filters: a bachelor’s listed, 30%+ of workers without one, 30%+ saying less is needed, 5,000+ openings a year. Teaching and clergy roles are left out because licensing boards or the employer set those rules. #### My recommendation. Ship the template and the test together. - Skills-first templates: List the role’s distinctive skills from O*NET, with the degree field optional by default. - An assessment per template: Templates alone repeat what most employers did. The test is what changes the screen. - Match the test to the skill: Management roles get a budget or staffing case. For technical roles, a hands-on task. Sales candidates do a role play, and compliance candidates write a short memo. - Launch with the ten: About 260,000 skills-first openings a year between them. - Show customers the proof: Report hires made without a degree and how long they stay, in each customer’s own data. #### Limits. What this data can’t tell you. - Older hires: Some workers without a degree were hired before the degree became the norm, so the gap may overstate new hiring. - An estimate of scale: Openings times a share sizes the opportunity. It doesn’t predict actual hires. - Validate the skills: O*NET ratings come from analysts and workers. Each customer should check assessments against its own top performers. #### What’s next. What I’d study once templates ship. - An A/B test on posts: Compare applicant pools and pass rates for the same role posted with and without the degree line. - Assessment validity: Check whether assessment scores predict 12-month ratings better than the degree did. - Adjacent jobs: Use O*NET skill overlap to find which current jobs feed each launch role, so recruiters know where to source. - Backslide alerts: Flag customers whose share of non-degree hires drops after launch, the pattern Burning Glass saw in 18% of firms. - Pay outcomes: Join BLS wage data to show what a skills-first hire earns against their last job. #### TL;DR - Studied: How many jobs ask for a degree their own workers don’t hold, across 177 bachelor’s-listed occupations. - Found: In bachelor’s-listed jobs, 3 in 10 workers have no bachelor’s, 12.7M people in all. - By the numbers: 12.7M Workers without the degree their job lists | 970K Skills-first openings a year | 0.48 BLS and O*NET agreement - The catch: Some workers without a degree were hired before it became the norm, so the gap may overstate new hiring. - Next: A/B test job posts with and without the degree line, and compare the applicant pools. ### New-hire ramp time case study Page: https://www.andrewlengly.com/ramp-time-casestudy Open New-hire ramp time: https://www.andrewlengly.com/studies/ramp-time.pdf Built with: Python, pandas, statsmodels, DuckDB, Jupyter, Census LEHD A data study of why new hires leave in their first two years and how long they take to reach full speed, modeled on 1,470 employees. It compares new-hire and settled-worker pay across California sectors and prices out a buddy program. Built with Python, pandas, statsmodels and DuckDB, it recommends a buddy for every new hire and a ramp time KPI. - The question: New hires take about a year to reach full speed, and many leave before they get there. What keeps them, and how should People Ops measure it? - What I did: Modeled early exits for 1,470 employees, compared new-hire and settled-worker pay across California sectors, and priced out a buddy program. - Where it stands: Done in November 2026. I recommend a buddy for every new hire and a ramp time KPI read at 30, 60 and 90 days. #### The question. A new hire takes a year to get up to speed. Many don’t stay that long. Gallup estimates new employees take around 12 months to reach full performance. Only 12% of employees strongly agree their company onboards people well. That year is expensive. I wanted to know who leaves during it and why. I also wanted a measure that lets a company see the problem early. Source: Gallup on onboarding, https://www.gallup.com/workplace/247076/Gallup_Perspective_on_Creating_an_Exceptional_Onboarding_Journey_for_Your_New_Employees.pdf #### How I did it. An inside view from HR records and an outside view from Census pay data. From records to a KPI: 1. Tenure curve. Attrition by years at company. 1,470 employees 2. Early exits. Logistic regression. First two years only. Controls: Job level, overtime, pay and role 3. Census QWI. New-hire vs settled pay. California sectors, 2025 4. Payback. Lost output while ramping. Buddy cost and break-even 5. KPI. Five measures. Read at 30, 60 and 90 days - early exit = leaving within the first two years at the company - new-hire pay ratio = average pay of new hires who stayed a full quarter / average pay of all settled workers - lost output = salary times the average gap to full output over a 12-month ramp that starts at 25% The ramp shape and every dollar input are assumptions you can edit in the study’s assumptions file. #### What the data shows. People leave most in their first two years. Attrition by years at the company: - Under 1 year: 36% - 1 year: 35% - 2 years: 21% - 3 years: 16% - 4 years: 17% - 5 years: 11% - 6 to 10 years: 12% IBM HR Analytics dataset, synthetic. 43% of everyone who left was in their first two years. First-two-year attrition, by years with the current manager: - Under 1 year (212): 35% - 1 year (26): 23% - 2 years (104): 20% Employee counts in parentheses. With controls, a manager of under a year meant 2.3 times the odds of an early exit (interval 1.2 to 4.1, p = 0.007). Training sessions showed no clear effect (p = 0.18). A new manager stands in for a working relationship that hasn’t formed yet. That’s the gap a buddy fills. Overtime had the largest effect in the model, with 6.3 times the odds of an early exit. #### The outside view. California’s new hires earn 60% of what settled workers make. New-hire pay as a share of settled pay, California, 2025: - Information: 79% - Construction: 77% - Accommodation and food: 75% - All private: 60% - Management of companies: 51% - Education: 49% Source: Census Quarterly Workforce Indicators, https://lehd.ces.census.gov/data/ New hires also skew younger and more junior, so the gap is a rough stand-in for ramp. It still shows how far a typical hire starts from full contribution. #### The buddy math. A buddy pays for itself if ramp shrinks by 18 days. Microsoft hires who said a buddy sped them up, by meetings in 90 days: - 1 meeting: 56% - 2 to 3: 73% - 4 to 8: 86% - 9 or more: 97% Source: Microsoft Workplace Insights, https://workplaceinsights.microsoft.com/?p=853 - $24.4K Lost output per hire while ramping - $1,230 Buddy cost per hire - 18 days Faster ramp to break even - $167K Net per 100 hires All four figures rest on assumptions. The model uses a $65,000 salary with output that starts at 25% and climbs over 12 months, and a buddy who gives 2 hours a week for 12 weeks. The net figure also assumes one month faster ramp and 3 points fewer first-year exits. #### The KPI. Five measures, read at 30, 60 and 90 days. - Manager readiness rating, 1 to 5 - Target: 4 or higher by day 90 - From: A short manager pulse in the HRIS - First work shipped on their own - Target: Shipped by day 60 - From: Manager pulse - Buddy meetings held - Target: 9 or more by day 90 - From: Calendar or a buddy check-in form - Knows who to ask - Target: 90% say yes by day 30 - From: New-hire pulse survey - Still employed - Target: 95% or more at day 90 - From: HRIS #### My recommendation. Give every new hire a buddy, and measure the ramp. - Assign before day one: Pick someone who reports to the same manager. Microsoft found they know the role best. - Nine meetings: Set a target of nine or more buddy meetings in the first 90 days. - Clear the buddy’s time: Budget about two hours a week for 12 weeks and take that work off their plate. - Report it monthly: Put the five ramp measures next to 90-day retention. - Watch overtime: It was the strongest predictor of early exits in the data. #### Limits. What this data can’t tell you. - Synthetic data: IBM built its HR dataset for teaching. The direction matches Microsoft’s and Gallup’s research. - Manager tenure: Years with a manager can also reflect reorganizations and manager turnover. - Self-reported: Microsoft’s buddy figures come from the new hires themselves. #### What’s next. How I’d prove the buddy effect. - Staggered rollout: Start buddies in some teams a quarter early and compare ramp and 90-day retention, a cleaner test than any model. - Meeting dose: Log buddy meetings and check whether Microsoft’s pattern holds in hard outcomes like manager ratings. - Overtime alert: Flag new hires working overtime in their first 90 days, since overtime was the biggest early-exit signal. - Cohort pay curves: Follow a QWI cohort over several quarters to see how fast new-hire pay actually catches up by sector. - Remote onboarding: Split the KPI by remote and on-site hires, where building a network works differently. #### TL;DR - Studied: Why so many new hires leave in their first two years, and what a buddy program is worth. - Found: 43% of everyone who left was in their first two years at the company. - By the numbers: 36% Attrition under 1 year at the company | 2.3 times Odds of an early exit with a manager of under a year | 6.3 times Odds of an early exit with overtime - The catch: All 1,470 employees come from a synthetic dataset IBM built for teaching. - Next: Start buddies in some teams a quarter early and compare ramp and 90-day retention. ### Internal mobility case study Page: https://www.andrewlengly.com/internal-mobility-casestudy Open Internal mobility: https://www.andrewlengly.com/studies/internal-mobility.pdf Built with: Python, pandas, statsmodels, DuckDB, Jupyter, BLS API A data study of whether people who move inside a company stay longer, modeled on attrition and time since promotion for 1,470 employees. It compares people who changed roles with people who did not and sizes the value for a 5,000-person company. Built with Python, pandas, statsmodels and DuckDB, it recommends a two-week internal posting window and skills-profile matching. - The question: Companies post a role outside before most of their own people even see it. What does that cost in people who leave? - What I did: Modeled attrition against time since promotion for 1,470 employees, compared people who changed roles with people who didn’t, and sized the value for a 5,000-person company. - Where it stands: Done in November 2026. I recommend a two-week internal posting window and skills-profile matching. #### The question. People leave for a new job they could have found down the hall. LinkedIn found employees lean toward outside jobs because those openings are easier to find and pay more. Gartner found only about half of employees know about openings at their own company. Workday’s talent marketplace sits on top of that gap. I wanted to know how much staying put in one role raises the risk of leaving, and what that risk is worth. Source: LinkedIn Global Talent Trends, https://business.linkedin.com/hire/global-talent-trends/archival/global-talent-trends-may-2023 #### How I did it. Two HR datasets for the pattern, government and LinkedIn data for the size. From records to a dollar figure: 1. Promotion model. Logistic regression. 1,470 employees. Controls: Tenure, time in role, level, overtime, pay and satisfaction 2. Movers vs stayers. 2+ years at the company. With and without controls 3. Exit reasons. 88 voluntary exits. Huebner HR Dataset 4. Sizing. JOLTS quit rate. LinkedIn retention gap - mover = someone whose time in their current role is shorter than their time at the company - kept by moves = extra internal moves times the gap between 75% and 56% two-year retention The IBM data can’t separate promotions from lateral moves, so a mover is anyone who changed roles. #### What the data shows. Each year without a promotion raises the odds of leaving by 15%. Chance a typical employee leaves, by years since promotion: - Just promoted: 6% - 1 year: 7% - 2 years: 8% - 3 years: 9% - 4 years: 10% - 5 years: 11% - 6 years: 13% - 8 years: 16% Everything else held at the median. Odds ratio 1.15 per year, interval 1.07 to 1.23, p < 0.001. IBM HR Analytics dataset, synthetic. Share who left, people with 2+ years at the company: - Same role since hire (IBM, 176): 19% - Changed roles (IBM, 1,079): 12% - No internal move (LinkedIn, at 2 years): 44% - Internal move (LinkedIn, at 2 years): 25% Source: LinkedIn Talent Connect 2022, https://news.linkedin.com/2022/october/talent-connect-2022 The mover gap in the IBM data is significant on its own (p = 0.02). With controls the odds ratio is 0.67 and the interval runs from 0.41 to 1.10, so that result alone can’t settle it. LinkedIn’s data across millions of profiles points the same way. In the Huebner data, 40 of 88 voluntary exits gave a reason an internal move can answer. Leaving for another position was the most common. #### Opportunity. Four more points of internal moves is worth about $1.4 million a year. - 1,330 Quits a year at 5,000 people - 200 Extra internal moves - 38 More people still there at 2 years - $1.39M Value per year of moves BLS puts the private-sector quit rate at 26.6% for 2025. Raising internal moves from 8% to 12% a year adds 200 moves, and LinkedIn’s 19-point gap keeps about 38 more of those people. At $4,700 per hire plus about $24,400 of lost output while a replacement ramps, that avoids about $1.1 million. Skipping part of the outside search adds about $0.3 million. A move often opens a backfill, so only 30% of search cost counts as saved. The move rates are assumptions. #### My recommendation. Make the internal candidate the default. - Two-week internal window: On by default for every requisition, with exceptions for urgent backfills and licenses nobody inside holds. - Tell matched people first: Notify them the day a role posts. About half of employees don’t know about openings at their own company. - Skills profiles: Employees fill them in and managers confirm, with fields for career interests and willingness to move teams. - Rank by risk: Sort matches on skill overlap, then surface people three or more years past their last promotion first. - No manager veto: Managers see an application after the employee applies and can’t block it. - Report the payoff: Show customers their internal fill rate and two-year retention of movers. #### Limits. What this data can’t tell you. - Synthetic data: Both HR datasets were built for teaching. LinkedIn’s and Bidwell’s research carries the outside evidence. - Who gets moved: People who move may already be stronger or more engaged. A real test compares matched employees. - Moves are mixed: The IBM data treats promotions and lateral moves as one group. #### What’s next. What I’d build on top of this. - A flight-risk list: Turn the promotion model into a monthly list of people past the three-year mark with a matching open role. - Promotion vs lateral: Split movers by type with real HRIS data to see which kind of move keeps people longer. - Manager hoarding: Measure which teams lose few people to internal moves and still see high outside quits. - Window test: Compare fill time and quality for roles posted with and without the two-week window. - Pay gap tracker: Compare internal move raises with outside offers, since bigger pay bumps pull people to outside jobs. #### TL;DR - Studied: Whether people who move inside a company stay longer, modeled on attrition for 1,470 employees. - Found: Each year without a promotion raises the odds of leaving by 15%. - By the numbers: 1.15 Odds ratio of leaving per year since promotion | 19% Left, same role since hire | 12% Left, changed roles - The catch: Both HR datasets are synthetic, built for teaching. LinkedIn’s research carries the outside evidence. - Next: Turn the promotion model into a monthly list of people past three years with a matching open role.