I build agent systems, LLM tooling, and production-grade sandboxes — the infrastructure that makes frontier models useful in practice.
My work sits at the intersection of reasoning architectures, tool use, and deterministic UI enforcement. I care about engineering patterns that scale, not prompt hacks.
Languages Python · TypeScript · JavaScript
AI/ML DeepSeek · OpenAI SDK · Qdrant · Vector RAG
Backend FastAPI · WebSocket · SSE · SQLite
Frontend React · Vite · Tailwind CSS
Infra Docker · Playwright · GitHub Actions
Single model, variable effort. One frontier model handles everything from greetings (thinking OFF) to system design (reasoning max). The scaffolding adapts, not the model.
Diversity via prompt variation. When temperature is locked in thinking mode, Best-of-N achieves diversity through approach angles, not parameter jitter.
Deterministic enforcement. Design systems, sandboxes, and agent guardrails should be code — not vibes.