Applied AI · Building AI systems that do real work.
AI becomes much more interesting when it can do the work, not just help you do it.
The hard part is everything around the model: understanding context, using tools, working across systems, reasoning through ambiguity, taking action, recovering from failure, and reliably finishing meaningful work.
That is what I build around.
- AI agents and workers with LangGraph and OpenAI Agents SDK
- Reliable agent systems with evals, guardrails, observability, durable execution, and human approval
- Knowledge systems with Postgres, pgvector, hybrid search, and retrieval
- Agent interfaces with Next.js, FastAPI, Vercel AI SDK, and streaming UX
- Voice agents with LiveKit and real-time orchestration
- Production systems where agents interact with real tools, data, and workflows
Moving AI from impressive demos to dependable systems.
A useful agent is not one that appears autonomous. It is one with clear boundaries, the right tools, measurable quality, reliable execution, and enough visibility to understand what happened when something goes wrong.
- Contributed to LangGraph Open Canvas
- Built AI agent systems for multiple startups in 2025
- Specialty: Multi-agent workflows, canvas-style UX, voice AI integration
- Canvas Callback - Canvas-style AI interface (80+ stars)
- LangGraph Voice Call Agent - Real-time call to a LangGraph agent over LiveKit
- OpenAI Agents Streaming API - Stream responses from openai-agents-python using fastapi
- postbot3000 - (280+ stars)
The next step is not simply making agents more capable.
It is making them capable of owning well-defined pieces of work.
That means better context, better tool use, better memory, better evaluation, better execution, and better interfaces between agents and the systems they operate.
Paperclip · LangGraph · OpenAI Agents SDK · Vercel AI SDK · RAG · pgvector · Python · FastAPI · SQLModel · PostgreSQL · Next.js · TypeScript · React · Docker
