What I build

I work across the full stack around the model. I build data pipelines and analytical services in Python, Postgres, Azure Data Lake, and DuckDB, and expose them over MCP so agents and people query the same source of truth. I build agent workflows with LangGraph and Deep Agents, hybrid retrieval with Pinecone and Cohere, and FastAPI and Next.js applications with streaming interfaces.

I make long-running agents behave like production software: durable state, human review at the right moments, tracing, evals, cancellation, and recovery. I also build agents that improve with use through long term semantic memory, deciding what they should remember, update, and forget.

How I work

I start with the work itself. The user, the workflow, and what a useful result actually looks like. If the process already exists, I watch it being done. People rely on checks, exceptions, and workarounds they no longer think to mention, and that's usually where the real spec is.

Then I draw the boundaries: what stays deterministic, where to add AI, and which decisions stay with people.

I build a small version that runs end to end, put it on real work early. What it gets wrong, and what people wish it could do, decides what I build next.

In practice, it's agents, data pipelines, MCP servers, evals, and the harness that connects them.