ai infrastructure · harnesses · agents · evals · graphs · inference engineering
i build ai systems that reason — not just retrieve. the work lives in the seams: harnesses that hold models accountable, agents with bounded agency, evals that catch failure modes before they ship, inference engineering that makes it run where it matters.
security roots. editorial eye. accuracy > speed.
| languages |
|
| ai / ml |
|
| local inference |
|
| infra / ops |
|
| frontend / 3d |
|
| project | what it does | stack |
|---|---|---|
| bahia-rosa · live | the city prints you — one photo in, a painted character portrait out, entirely in the browser: the segmentation, the palette, the compositor and the editor all run on the visitor's machine, and the photograph never leaves the page | TypeScript React 19 MediaPipe Unlayer |
| forge · live | a commander agent with verified, cumulative learning — a typed execution graph, a memory fabric instead of raw transcripts, and capabilities promoted only through an A/B gate | Python Multi-agent Evals |
| clusterbreak · live | build a rig, run a local model, break it on purpose — simulated tokens/s, TTFT and VRAM with causal postmortems, then real CloudFormation and STS provisioning on a g5 | TypeScript AWS CloudFormation |
| razorpay-agent · live | a merchant-side agent that proposes offers to an autonomous buyer over ACP and settles them on Razorpay — every action bounded, gated and audited | Python ACP Razorpay |
| sword private · in progress | a gpu-accelerated, agent-friendly terminal emulator — the VT parser is an enum dispatch over states, not vte |
Rust GPU VT |
