Projects

Backscroll is an agent that scrolls your feeds. It scrapes my twitter and instagrams feeds and imports as much historical likes and bookmarks to send to a llm to compile a feed that you really want to see.

How many words can you stuff in the slit that CC, Codex, and nearly every chat interface give you? You can steer the session with every subsequent message but when you steer a car you want tactile, physical, and nearly intimate way of interfacing with the interface that gets you from A to B. This is an experiment in thinking about interactive interfaces.

Geospot was an app that used image embeddings to guess where you were in the world from a single image and no GPS. It was contrastively trained between GPS coordinate pairs and images from 11.7 million street view images.

I scraped 200,000 tweets, around 10,000 hours of TikTok videos, and hundreds of hours of standup sketches to post-train a 1 trillion parameter model. I used rubric RL to have 8 judges score rollouts from Kimi K2 based on the rubric and upsampled the best jokes that satisfied the rubric.

The idea here was to build a search engine that was dense enough with enough variety of ideas that even if we can't service your specific search query we can give you something pretty reasonably interesting to read; to fall into a rabbithole of human curated web of links.

I built Cuaview as a native harness around Claude Opus 4.5 to see what a strong computer-use model could do one-shot on real tasks. Through that loop it could play League of Legends badly but coherently, play Minecraft, book flights and hotels pretty well if you asked nicely, and do taxes end-to-end.

Pingpong is a realtime analytics engine for sites I run. Cloudflare Durable Objects employs an actor-architecture perfect for this since: every site opens up a websocket to each site's DO we can handle multiple tabs, live precise, and a DB directly in each worker. Due to the nature of DOs we need to have a "parent" DO that rolls all this up to serve into this dashboard which is what you see.

I built the #1 ChatGPT plugin that lets you ask questions about any PDF you upload -- this was made before you could just upload a PDF to ChatGPT which is why it was so popular.

nanoEBM is a tiny, from-scratch, ~10M-param char Transformer (67 vocab, 6 layers, 384 dim, 6 heads) with a linear energy head. Its based off @karpathy's minGPT with an additional 400 lines for our EBM.

AI-native web OS for writing and editing spreadsheets, code, and docs with modular support for specific verticals. Built for a multi-billion dollar company, pivoted since it felt mediocre.

inspired by SDPO, i made continualcode -- a minimal claude code that learns from your corrections in real-time, built on tinker. when you deny a diff, the model uses your correction as context to teach itself, takes a gradient step on LoRA, and retries with updated weights.

for the first time you get persistent GPU/CPU sandboxes via Modal/CF backed by Durable Objects to stream logs live, native preview URLs, and attach a real shell

made a site to run deepseek r1 for free -- locally in your browser, no downloads, no servers purely using WebGPU

Sudo streams and renders interfaces generated just-in-time.

Bountylist is a small organic list of "bounties" which are 1-4 week projects that you want a talented builder, anonymous anime pfp tinkerer, or some student who's automated their homework with GPT-4 to complete.

Search all of Ye’s tweets and lyrics from the archives using an ultra fast vector database.

Now you can easily ask to graph, answer questions, and you'll get a beautifully rendered animation explaining the concept

By the end of the summer I built a sensor-fusion system for a RC car to be able to navigate a track with a neural network trained end-to-end.

I spent my first all-nighter and my first time having gorilla glue stuck on almost all my fingers to build a bridge that weighed under 15g and won the regional competition.