About
Building software with agents, deliberately.
AI coding agents are a core part of how I build. What matters to me is not how much they can produce, but the structure around them: clear rules, small verified steps, and a human who stays in charge of direction.
How I work
- Clear rules up front. Repo-level
CLAUDE.mdfiles and path-scoped rules give every agent the same conventions before it touches code. - Lessons that stick. Mistakes and incidents become written rules in a tiered memory, so the same problem is not solved twice.
- Isolation and verification. Each agent works in its own worktree, and important changes are independently reviewed before they land.
- Humans decide. Larger efforts are broken into task trees with explicit approval points. Agents propose; I approve.
- Sustainable pace. I invest in tooling and process so the work stays calm and repeatable rather than depending on sprints.
What it has produced
I built the whole stack of a fintech and lead its agent-assisted delivery. I have also built a Rust trading runtime and a quant research platform largely through agent operation.
Quality comes from enforced structure:
- CI with pinned actions and health-gated deploys, so a bad release stops itself.
- Written postmortems and runbooks that feed back into the agents' rules.
Open source
idea-vault is a local-first ideation tool in a single Rust binary.