Work
Production agents, in the field.
How do you take LangGraph agents to production? How do you detect documentation drift with an agent? How do you benchmark local models against production traces? Selected systems — teasers where IP requires it.
Production LangGraph agents
CDP AI Support System
Leading the architecture and build of the AI Support System for Coinbase Developer Platform — multi-turn agents with RAG, checkpointing, and observability. Covered in our LangChain Interrupt talk.
Agents as services
Coinbase AI Tiger Team
Cross-functional tiger team that set enterprise standards for code-first agentic systems — LangGraph patterns, auditability, and production hosting so agents ship like services inside Coinbase.
Documentation drift · Claude Agent SDK
Docs-quality agent
Production agent over the Claude Agent SDK that runs developer documentation in a secure sandbox, follows the guides to detect docs drift, and raises Linear issues plus GitHub PRs. High-level teaser — internals stay internal.
Public write-up forthcoming · related essays
Local LLM benchmark · production traces
Local↔production LLM evals
Ongoing experiments on Mac Mini and NVIDIA DGX Spark hardware: benchmarking open-source models against production LangSmith traces to see what holds up outside the cloud APIs.