Writing
Notes from the work. Real numbers, real screwups, real fixes.
How the systems I build actually got built: who needed them, what broke on the way, and the numbers that say whether it worked.
Posts ranked by how much the project behind each one moved the company.
How I turned a patient table into the practice's patient hub
The CDF dashboard started as one row per patient with every document one click away, and since September it is where the practice's other software asks who a patient is. This is how a read only Denticon mirror, a small registry API and measured record links got it there, including the green job whose eighteen calls all failed.
How I turned a hand built scheduling workbook into a live capacity planner
Our operations lead planned doctors, assistants and coordinators across three offices in an Excel workbook of pasted week blocks and fixed ranges. I replaced it with one forward view of demand, capacity and capability that reads Denticon every minute and never writes back. This is how it is built, what live data broke in the first day, and what is still waiting on answers.
How I moved visit plans out of Word and into the estimate
Coordinators used to type every patient's plan of visits into Word from a master document, then renumber and redate it by hand on every change. Now the estimate picks one of the practice's own transcribed masters, a three pane builder edits it, and Word and PDF come out the other end. Two packages in their own repo, vendored into the estimator with no registry and no token.
How I automated a finance forecast without letting anything touch the math
Every morning the VP of Finance pasted two lab exports into his Excel forecast, recalculated, and sent it by hand. I rebuilt it as a daily pipeline that ports his model exactly, refuses to tune itself, keeps Gemini away from every number, and never holds a credential that can read his inbox.
How I joined five marketing logins into one warehouse a model can read
Our marketing numbers lived in five separate logins, and no connector on the market could follow a click through a phone call and a CRM lead to a booked patient. I built a BigQuery warehouse that does that join every morning, describes itself for the language models that query it, and grades those models every week, including the first run that failed its own gate.
AI has the memory of a goldfish with a film degree
At the GDG Brooklyn hackathon at NYU Tandon we looked at a four hour timer and decided to build an entire AI filmmaking studio. Ask an image model for the same character twice and you get siblings. This is how we made a character stay one person across a whole film, and put the proof on screen as a number.
How I gave a three person insurance team 70 hours back every week
A three person insurance team at a multi office dental group spent about 70 hours a week cross checking three exports by hand to confirm every patient seen got documented and billed correctly. I built a deterministic engine that does the reading for them and hands back only the flagged patients. Shipped, live, and HIPAA safe by construction.
Teaching an AI to read the whole patient and lean toward a treatment, before the doctor does
The Centralized Diagnostic Filter is HYBRIDGE's diagnostic operating system. AI reads every input a complex case generates, surfaces the findings, scores future tooth-loss risk, and leans toward a treatment direction, all before the consult. Then the doctor validates every finding. The intake shipped and is live in production, the AI panorex read is real and running under a written ruleset extracted from the founder, and his review of the first outputs has already revised that ruleset once. The validation loop is not a diagram anymore.
Building an enterprise search that knows when to say 'I don't know'
Hybridge's knowledge is scattered across drives, decks, and people's heads. I'm building an internal search that answers questions in plain language, with citations, and that refuses to guess when it isn't sure. This is how I'm thinking about a RAG you can actually trust, built one stage at a time.
How I rebuilt a brittle call grader into a coaching platform
A prototype scored every front-desk call out of 60 and emailed a link to a Google Doc. I rebuilt it into a coaching platform: a verbatim quote behind every score, patient safety as a loud flag, the math kept out of the model, and a dashboard a manager actually opens. Here is the whole story, including the bugs my tests could not catch.
I built an open-source app to run my job hunt
JobPilot finds fresh roles across 160+ company boards, scores each against my profile, writes a tailored resume and cover letter per match, and tracks every reply to offer. It's now open-source, and it ended up touching all four things I do: data, ML, software, and shipping.
How I taught the consult grader to write report cards
We already score every patient consult. Now an AI reads all those scores and writes each doctor a coaching report card, every Monday at 8 AM, with nobody clicking anything. Plus the five bugs I squashed on the way.
How I built a job-search autopilot in 5 hours
Refreshing LinkedIn for jobs that were posted three weeks ago is not a strategy. So I built JobPilot: 7 job sources, 4 runs a day, 4 tailored resumes, one dashboard, for under $5 a month.
How I caught my pipeline lying to me (and made it confess)
Runs showed ✅ done while emails never sent, rows never appeared, and whole meetings just vanished. Nine kinds of silent failure, all hidden. The fix was making the dashboard stop lying.
How I saved our accountant 400 hours a year
Every week our accountant hand-copied numbers from five different systems. A dozen small Python scripts do it for him now, built the way he actually works.
How making movies made me a better engineer
People keep telling me films and engineering are separate. They're not. Here's what ten years of making movies did to the way I build.
How I built an AI that grades every patient consult
A Zoom call comes in. My tool finds the doctor, grades the consult against the CEO's 7-point checklist, and emails back a color-coded report, without making things up.
How I rebuilt a 10-year-old pricing tool in a month
A 10-year-old pricing tool, rebuilt in about a month around one rule: a saved price never changes, so the 6-month price promise holds.
How I built a clinic dashboard people actually trust
Two clinics, three messy boards, and a patient-to-lead match that was only 49% right. The fix wasn't a clever algorithm. It was reading the data properly.
How I replaced a $98K quote with a $50-a-month tool
A vendor wanted $98K up front and $26K a year to flag bad dental CT scans. I built it in-house on Cloud Run for under $50 a month, with one painful lesson.
How I trained Llama to argue more convincingly
I trained Llama 3.1 to write better comebacks in online debates, using a reward model and RLHF. Here's what actually moved the needle, and what didn't.
How I built the boring parts of an ML system on purpose
A loan-scoring service where I built the unglamorous parts on purpose: training in containers, model history, quality gates, automatic retraining. The gates mattered most.
How we went from 42nd to 1st in a class AI contest
A CIFAR-10 competition across 55 teams, a hard 5M-parameter cap, and two weeks. Tuning got us stuck at 92%. The thing that actually won it was almost embarrassingly simple.
How I built a recommender for 22 million records
I built a recommendation system that handles 22 million records: one fast step to narrow the field, one smart step to rank it. It beat 'just show what's popular' by 20%.
How I caught fake news and read the mood of an election
I built two models for the 2020 election: one to flag fake news, one to read how people felt across 1.8 million tweets. The modeling was the easy part.
No posts in this one yet.