// experiments

Going deeper

When I want to know if I can build something at a production standard, I build it myself, end to end, and hold myself to the same bar I'd use shipping to real users. These are the deepest dives — where I go past "good enough" to understand a system all the way down.

Built solo, to a production bar. These aren't tutorials or toy demos. They're fully-featured systems (guardrails, cost tracking, eval harnesses) that exist because I wanted to understand something completely, not because a customer asked for it. Some source lives in private repositories.

Turn-by-Turn Agent

A tool-using AI agent framework, built from first principles

A production-quality TypeScript monorepo that implements a complete turn-by-turn (ReAct) agent SDK from scratch. The aim is not to replace off-the-shelf frameworks but to deeply understand and document how a capable agent is built — every non-obvious design choice is captured in an architecture decision record. It ships with a working customer-support ticket-triage agent that wires all of the packages together. It's the clearest proof I have that I can build real infrastructure on my own instead of just using someone else's SDK.

Core ReAct turn loop with guardrails, checkpointing, and per-run cost tracking
Pluggable model providers (Anthropic, OpenAI) with automatic failover and error classification
A tool registry with a middleware pipeline and human approval gates
First-class Model Context Protocol (MCP) client with schema conversion
Context-compression memory and a full eval harness — recorder, replayer, and exact/contains/LLM judges

TypeScriptpnpm monorepoAnthropicOpenAIMCPBiomeVitest

Code