Retrieval that admits when it doesn't know
Grounding, citation, and a refusal path. An answer with no source is a bug, not a feature.
Full-stack & AI Engineer
I build systems that think and interfaces that don't get in the way.
Full-stack from the metal up — data plane, model, and the pixels in front of it.
Most of my work is deciding what not to ship.
Grounding, citation, and a refusal path. An answer with no source is a bug, not a feature.
A scored set beats an opinion. Every prompt change ships against a suite that can veto it.
Narrow tools, explicit state, hard budgets. Autonomy is earned per-capability, not granted wholesale.
Streaming, caching, and model tiering are UX, not infrastructure trivia.
Agentic workflow platform. Built the orchestration layer, the eval harness that gates every release, and the interface that makes a non-deterministic system feel accountable.
Hybrid retrieval over a few million documents. Reranking, citation spans, and a cache that cut median latency to a third without touching recall.
A collaborative editor with optimistic state and conflict resolution that never blocks a keystroke. Sixty frames a second on a five-year-old laptop.
260,000 particles simulated on the GPU every frame, morphing between five attractor states. WebGPU compute via vgpu, with a 2D fallback that degrades quietly.
Let’s makesomething good.