Most AI work fails not at the model but at the engineering around it: data, evaluation, cost control, and shipping something that holds up in production. That is where my background helps, since I have spent years building reliable data pipelines and full-stack systems.
I can take an AI feature or agent from idea to a shipped, evaluated, cost-aware product, and I am honest about where AI helps and where it quietly adds risk.
Most recent: CodeWisp (YC W26)
At CodeWisp, an AI game-generation platform, I built the generation pipeline as a chain of agentic stages, working directly with the founder in twice-weekly product calls.
- Sequential agent stages: routing, research, build, art, audio, quality gate and publish.
- Multi-model routing across Claude Opus, Sonnet and Haiku, matched to task complexity and kept under a fixed $6 per-run cost cap.
- An LLM-as-judge eval stage that checks sprite consistency, level passability, win states and mechanics, retries on failure and logs findings to improve the system prompts.
- AI code generation constrained to a Lua sandbox to cut hallucinated code.
- A deterministic 2D/3D runtime at 60 fps with hash-based verification and per-game engine versioning.
Before that I built Codepost, an AI code security scanner and PR-triage bot, and prototyped an AML alert-triage copilot and a helpdesk auto-QA tool. Read the CodeWisp case →
What I bring
- AI features and agents wired into real products, not demos.
- Evaluation, cost control, and reliability around LLMs: LLM-as-judge evals, multi-model routing, sandboxed code execution, RAG with embeddings and vector search.
- Working directly with founders and customers to scope, demo and iterate.
- Full-stack engineering foundation in Go, TypeScript, Node, and React.
Proof, not promises
Real work, written up plainly: