I build agent platforms, and the working discipline that makes them safe to hand real work.
pg-ai-stewards keeps agent state, memory and pipelines as Postgres rows. Code agents run sandboxed, a critic stage reviews their work, and a human gate sits in front of every merge.
chillacks is a room where Claude Code sessions talk to each other, so agents on several machines can hand work to one another directly.
lodestar maps many repos into one cross-service code graph, so a person or a model can pull in only the part of a system they need.
loom drives Claude Code, Codex, Copilot CLI and other agent CLIs as long-lived workers behind one Go interface.
working-with-ai is the method under all of it, packaged as a Claude Code plugin, and ai-jumpstart is the same method for any model. Most of its rules were paid for by a specific failure, and each one says which.
scripture-book is Beyond the Prompt, a book manuscript on where the method came from. I spent months learning to study scripture with an AI honestly, and the patterns of creation, covenant, stewardship and refinement turned out to be the same ones that make human and AI collaboration work.
multiplayer-ai carries the method from one seat to many: one person running a room of agents across a fleet of machines, and the norms that keep that room from drowning in its own messages. Its chapters land only once the practice behind them has run.
On inference, I have 24 merged pull requests in syv-ai/HyperQwen, which runs Qwen3.8-27B on a single RTX 3090, from speculative-decoding kernels to the ports onto vLLM 0.29 and 0.30. I also have an upstream fix merged in sgl-project/SpecForge.
Go, TypeScript, Python. Remote, from Missouri.