Long-form pieces on multi-agent orchestration, document intelligence pipelines, and the infrastructure work behind AI systems that actually reach production.
TanStack Start, Convex, and Clerk — plus the specific architectural patterns from Theo Browne’s open-source Lawn repo that make them work together.
Rebuilding MIT and Google Brain’s multi-agent debate paper as a chat product, end to end on Convex — sub-threads per model, two parallel rounds, a hidden synthesis prompt, and a Zod-validated summary.