Foreman
NewAsk an AI to watch security footage for hazards and it finds them everywhere, because that is what you asked about.
All work!
Six pieces — the ones I would defend in review.
A lot of company software has no way in except the screen. To get data out, something has to click through it the way a person would.
Proveit!
Builds the check,not just the thing.Based in New York.
lmost everything I make has the same shape. Something produces an answer — a model, a spreadsheet, a payment provider — and I build the part that works out whether to trust it.
I learned that on unglamorous data. For a year at Accenture I worked on commercial insurance records that arrived from brokers in whatever shape they felt like sending, and once found a $4 million property that had been entered as $40 million. Nobody had noticed. You cannot read a hundred thousand rows by hand, so you write the checks that find the row worth reading.
Everything since has been that same move pointed somewhere new: video, browsers, payments, city records. I graduated from NYU in May 2026 with a master's in Management of Technology, and before that took a computer science degree I still use daily.
Building things that ship
Years
Live, taking real payments
Countries
Hand-labelled to test my own work
Clips
Written, reviewed, in production
Lines
It started with a spreadsheet that lied.
A $4 million building had been keyed into an insurance record as $40 million. It had been sitting there for a while. Nobody had noticed, because nobody reads a hundred thousand rows — they read the summary, and the summary was wrong in a way that looked completely normal.
What fixed it was not attention. It was a set of cross-field checks that could say: this number disagrees with that number, go and look. I spent a year building those, and watching people trust the output more once they could see what it had caught.
Models are wrong in exactly the same way. Confidently, plausibly, in the shape you were expecting. So I build the same thing for them — the eval, the second pass, the number you can go and check yourself. That is the whole job, and it is why every project here ships with what it cost as well as what it did.
— Y.N., New York
The record
Accenture
- Commercial-insurance data covering 1,000+ properties, inside a team of about ten.
- Built the cross-field checks that caught a $4 million property entered as $40 million, and wrote the brief that got it corrected.
- Took AI-assisted validation from 40% to 95% adoption in eight weeks, using a log of what the checks caught to win people over.
- Cut correction cycles by 20 percentage points across compliance, actuarial and data engineering.
Amoga
- Owned a CRM product end to end at a startup: research, stories, backlog, sprints, and the go-to-market around it.
- Grew web traffic 7% and lead conversion 10% through outbound campaigns and funnel dashboards in Superset and Google Analytics.
- Wrote the buyer-facing copy and sales kits, because there was nobody else to write them.
- New York University — MS, Management of Technology
- Ramaiah Institute of Technology — BE, Computer Science
The archive
Team of four; I led the analysis, not the model. Found that stockouts and healthy stock had near-identical inventory buffers, which broke the obvious threshold rule everyone starts with.
57 engineered features predicting which drivers take a coupon, turned into a recommendation about which amenities a highway should actually build.
Twenty thousand app reviews sorted into themes, then mapped onto what a product team should fix first.
Recency, frequency and spend, in SQL, visualised in Tableau. The plainest possible version of turning a table into a decision.
Measure the thing you built
A demo tells you something can work once. A measurement tells you how often it works, and what it costs when it does. I would rather hand you the second one, even when the number is worse than the story.
All workThe model read a sign and called it an accident.
Ask a model to watch footage for accidents and it will find accidents, because finding them is what you asked about. Mine once reported a pedestrian walking in front of a moving forklift, at 95% confidence. A second model agreed and marked it high severity.
The frame was a black title card. Printed on it, in white letters, were the words PASSING IN FRONT OF A FORKLIFT. It had read the words and filed them as something it had seen.
Printed text is everywhere in real footage — slates, timestamps, captions burned into the picture. One extra check in the same call removed that whole category of mistake at no added cost. Finding things like this is most of the job, and it is why I build the measurement before I trust the demo.
The back page
What I'm into, off the clock: the things I would talk about for an hour if you let me. One of them got out of hand and is still running.
- Hong Kong cinema
- Stephen Chow first, then the action films, then Wong Kar-wai when I want to feel something. Mo lei tau — the nonsense-comedy style Chow works in — is the thing I will talk your ear off about. King of Comedy was hard enough to find that I ended up watching it on YouTube with subtitles from somewhere else entirely.
- Films the ratings got wrong
- A film almost everyone scored badly and I think is good. That disagreement is roughly why I do not keep a Letterboxd: I would rather come to something cold than through someone else's score.
- Chess
- I never learned any theory. I have opened with the London 2,669 times and worked out everything after move four as it happened, which has got me this far and will presumably stop working at some point.
- Going places cheaply
- I like going places and I graduated broke, which is a bad combination. So I wrote something to watch flight routes out of New York and tell me when a fare drops far enough below its own baseline to be worth taking. 214,882 readings since July. I have booked two of them.
