Neil Gilani

Cameras that spot sick livestock days before a person would.

Cattle hide illness when someone walks past, so a stockman checking the pen sees animals that look fine. Vitals watches when nobody is there. It detects and re-identifies each animal, tracks how long it feeds, moves and lies down, and compares that to the same animal's own 14-day baseline. When several signals drop together, it raises an alert.

I'm working on a solar-powered version that runs the model on the camera itself, with a target of about half a watt on average.

A ski and snowboard game that runs in the browser.

There are eight mountains, each with its own named trails and lighting, from a Japanese powder bowl at dusk to a Norwegian run that ends at the sea. The physics handle carving and skidding, jumps, spins, flips, grabs and rail grinds, and a crash turns the rider into a ragdoll. Tricks feed combos, XP and unlockable gear.

I build it with a friend. We split big features across parallel Claude sessions, one per area in its own git worktree, and one integration session merges them and runs the tests.

A Mac home screen and accessibility layer for people who can only move their eyes, such as people with ALS.

You pick a goal with a few selections by gaze, blink or a single switch. An AI agent carries out the steps across apps, and anything consequential, like placing an order or sending a message, needs your approval first. That rule is enforced in code: each approval is tied to a hash of the exact action and can only be used once. It started at a UCSC AI hackathon, and I'm now working on launching it as a nonprofit.

I built it by running nine Claude Code agents in parallel, each owning one part of the system, with one acting as lead and merging pull requests.

Hold a paper sketch up to your webcam and get a working app.

It reads the drawing with OpenCV, Tesseract and a vision model, turns it into a typed description of the interface, and renders it live. You can drag, resize and retype anything on the canvas. "Build real app" goes further and produces a multi-screen Flutter app with login, a backend and live data that you can download.

Upload a photo of anything and get a brick model of it, with step-by-step instructions and a parts list.

Both AI models (depth estimation and subject segmentation) run in the visitor's browser, so the site costs nothing to run, needs no account, and has no build limit. The models shape the subject into a relief, then a brick engine places bricks, slopes and plates, checks that every piece is connected, and splits the build into steps.

Memory for AI models that keeps learning after deployment, without retraining.

Hebb attaches a writable memory to a model whose weights stay frozen. New knowledge is written into memory pages by gradient descent on those pages alone, and the model reads them back through attention at every layer. I wrote a paper on the main finding: what makes the model retain old skills is how memory is allocated, not how it's written.

Photograph a skin, eye, mouth or nail concern and find out how soon to get it seen, and by whom.

It's a triage tool, not a diagnosis, and it never tells anyone they're fine. Fixed safety rules sit on top of the model and flag emergencies, like chemicals in the eye, no matter what the model says.