making stuff around harnesses, evals and inference. currently into Apple Silicon and inference engg. exploring harnesses that solve hard problems like computer use, kernel generation etc.
occasionally i work on fpga's and risc-v
| project | about |
|---|---|
| qwendoom | the real doom, running inside a language model. nothing is trained: every token the model writes is one step of a small computer, and the game lives in the model's own memory. made it 47x faster by skipping parts of the model that can't change the answer. it has drawn its first 41 frames, every pixel matching the real machine. |
| bytesized | a real llm (smollm2-135m) run by a 2,946-byte c program, small enough to fit in one qr code. writes the exact same words as hugging face, token for token. |
| carmen | an ai writes gpu kernels for apple silicon, and a judge it can't fool decides which ones are real. fused 8 mlx kernels into 2 per layer of qwen 2.5: 1.03x faster decode, identical output. the judge caught 75/75 planted bugs; a standard one-shape check caught 40. |
| flowgap | finds what callers wanted that your voice agent can't handle, straight from raw call transcripts, and drafts the missing flow nodes. in 298 real calls, 49% asked for something the flow couldn't do. |
| ml / ai | |
| backend | |
| cloud / infra |  |
| databases |