Writing
Six case studies, then 37 posts and 2 guides.
A dental implant practice with three offices and its own lab runs on 45 live services I built. Here is the map, every service on it, and the patterns that keep them running.
45 services live on Google Cloud across 14 projects, checked on Oct 29, 2026
Our coordinators had four tabs and four ids for one patient. I built a read only Denticon mirror, a registry other apps share, and record links that prove a match before they write.
453 of 453 blind test against bookings: every time the record linker picked a chart, it was the right one
A vendor quoted a six figure build to flag bad dental CT scans. I built the checker in house in 37 days. It runs on CPU. Our testing team checks it on real studies.
$62 on Google Cloud from October 1 to 28, measured, against a six figure vendor quote plus a monthly fee
Consult reviews were written by hand against a seven criterion Word doc. Now a Zoom webhook starts a Gemini grader, and in October the median report landed 2.7 minutes after the transcript.
471 consults scored live from May 19 to Oct 28 across Rochester, Buffalo and Syracuse
JobPilot watched 163 company boards, scored what survived the filters with Gemini, rewrote each resume against a strict judge, and drafted every email for me to send. Here is the whole system.
1,008 runs of the pipeline from June 10 to October 8, every 30 minutes at first, then twice each morning
Hybridge staff could ask a question and get an answer cited to the exact passage, filtered by who may see what. Built in a month, paused since October.
24 days from the first commit to real models in production, with sign in, permissions and cited answers
Constellation
A forward deployed engineer works beside the people who have a problem and ships the fix into their systems. Here is that job at a three office dental practice, each example linked.
The patient site and the lab site, rebuilt in the premium lab look by parallel page agents with a marketing review, checked at 1440 and 390 and previewing on Cloud Run.
A dental implant practice with three offices and its own lab runs on 45 live services I built. Here is the map, every service on it, and the patterns that keep them running.
Six measured Google Cloud projects for a dental practice cost $131.71 from October 1 to 28. Two always on databases are two of the three biggest lines. Here is the table and what drives it.
Doctors who scanned, then sent the Rx, ended up with two lab cases and got scared when we cancelled one. I built a portal where the Rx joins the scan's case. Early, on sample data.
Asking Claude about our marketing data took the Google Cloud SDK and a fresh login most days. I built a connector: sign in with Google once, see only the data an owner ticked.
Change requests for our apps arrived in chats and in passing. I built one request screen six apps share, and a relay that trusts each app only to say who is asking.
Staff needed one place to open every app, leaders wanted to see what was being built, and change requests got lost in chats. I built one board and one request list six apps share.
I built the lead forms for our practice websites around one rule: no inquiry is ever lost. Each one is saved in our own Google Cloud first, then put on the Monday board.
Our coordinators had four tabs and four ids for one patient. I built a read only Denticon mirror, a registry other apps share, and record links that prove a match before they write.
Our operations lead planned three offices in a workbook of pasted week blocks. I replaced it with a planner that reads our Denticon mirror every minute and cannot write back. Seventeen days to production.
Coordinators retyped every plan of visits into Word. Now the estimate picks one of eleven transcribed practice masters and a builder edits it. The engine ships as a committed snapshot, with no registry.
The VP of Finance pasted two lab exports into a 14 tab workbook every morning and sent the forecast by hand. Now a pipeline builds the same forecast by 6:25 each weekday.
Our marketing numbers lived in five logins, and nothing joined an ad click to a booked patient. I built a BigQuery warehouse that makes that join every morning and mails it at 6:30.
Two apps needed the appointment book, and only one service may hold the Denticon keys. I built a small read only API over our mirror, with a secret per caller and strict offices.
Four Hybridge services need Denticon, and only one may hold its keys. I built the sync that keeps our copy fresh: webhooks first, a nightly poll behind them, and a check that notices silence.
At a GDG Brooklyn hackathon at NYU Tandon, four of us built an AI filmmaking studio in four hours. The hard part was keeping one character the same person across a whole film.
Our insurance team read the schedule, ledger and notes against each other by hand every day. I built an engine that does the reading and returns only the patients who disagree.
Intake and experience surveys with Gemini writing the doctor's page and a panorex read under the founder's own rules. Every call falls back through newer models, and clinic emails wait for both reports.
Hybridge staff could ask a question and get an answer cited to the exact passage, filtered by who may see what. Built in a month, paused since October.
The old grader sent a number out of 60. Mine had graded 690 real calls by October 29, a quote behind each score. October's median grade came 14 minutes after the call ended.
Every patient survey becomes a summary and a doctor report, emailed to the patient and the office. I built one renderer that three callers share, records each step, and finishes the job on retry.
JobPilot watched 163 company boards, scored what survived the filters with Gemini, rewrote each resume against a strict judge, and drafted every email for me to send. Here is the whole system.
Every consult already had a score. Now Gemini reads them all and writes each doctor a report card, every Monday at 8 AM and on the first weekday of each month.
I built JobPilot to pull fresh jobs from the source, drop the roles I could never take, and score and tailor the rest while I slept.
Runs went green while emails never sent. I moved the run row to the top, made every failure retry or turn amber, and added logs anyone can read.
Our controller ran nine Python scripts by hand, with 26 terminal prompts between them. I put them behind one page that asks every question before the run, in one evening.
People keep treating films and engineering as separate crafts. What making movies since 2014 did to the way I build software, from blocking a scene to cutting a feature.
Consult reviews were written by hand against a seven criterion Word doc. Now a Zoom webhook starts a Gemini grader, and in October the median report landed 2.7 minutes after the transcript.
Coordinators quoted from one shared price list with no office prices or tiers. The estimator I built writes the price onto every saved line, and shipped 14.5 hours after its first commit.
Our leadership numbers came from Monday boards joined by name. Fixing the join linked almost every patient to a lead. Then I moved it to Cloud Run, defined every metric once, and made cash flow daily.
A vendor quoted a six figure build to flag bad dental CT scans. I built the checker in house in 37 days. It runs on CPU. Our testing team checks it on real studies.
Research at NYU. Replies that earned a ChangeMyView delta became training pairs, a reward model learned to judge them, and GRPO let a fine tuned Llama 3 8B practice against it.
A three person NYU class project that scores loan default risk from the public LendingClub book. I built the data pipeline and the tests, and made training and serving share one transform.
A class contest: CIFAR-10, a ResNet, a hard cap of 5 million parameters. Flip and crop topped out at 93.76%. Augmentation, which adds zero parameters, took us to 97.12%.
A three person NYU capstone on 33.8 million MovieLens ratings: Spark ALS for recommendations, and my MinHash notebook that found movie twins after Spark's similarity join kept losing workers.
A three person NYU project on the 2020 election: a fine tuned Llama 3 labels tweets true, false or opinion, RoBERTa reads their mood, and I turned the results into state maps.
No posts match. Try another tag, year or search.
Teaching
A visual, first principles guide to retrieval: one mental model that runs from a bare baseline to what the best companies ship in production.
Read it
A working guide to building AI systems that ship: retrieval, agents, evaluation, and everything around them.
Read it
