AI systems builder based in San Francisco. My production background spans multimodal evaluation, high-volume data systems, workflow automation, and enterprise agents. My current public focus is the infrastructure underneath reliable coding agents: memory, delegation, and evidence-driven development.

Ten bug fixes and two small features in projects other people depend on, each merged by that project's maintainers. I use AI coding agents to find candidates and draft fixes, then reproduce each bug, verify the fix and its regression test locally, and see the pull request through review.

• axum (Rust web framework): the Allow header no longer lists HEAD twice after a get route is merged with a separate head route; four review comments addressed the same day. #3836
• SQLx (Rust SQL toolkit): pre-1970 datetimes stored as SQLite REAL (Julian day) values no longer decode up to two seconds off. #4340
• AnyIO (Python async I/O): the asyncio CapacityLimiter no longer grants more tokens than it has when total_tokens is raised while it is over-subscribed. #1223
• goose (open-source AI agent): saving a custom model that is not in the provider's list as the default for a new chat no longer fails with Invalid params, a regression three users reported. #10438
• Fastify (Node.js web framework): content-type parsers registered with a global or sticky RegExp no longer miss matches because of a stale lastIndex; approved by two maintainers. #6846
• Zod (TypeScript schema validation): .catch() callbacks receive the original input instead of the coerced value, as the documentation describes. #6192
• TanStack Query (data fetching): combine results that are falsy (0, false, "", null) are memoized instead of being recomputed. #11065

Also merged: a narrow floating-point edge case in Tenacity's wait_exponential (#656); the maintainers' proposed fix for contradictory PyArrow constraints that made feast[flink] uninstallable in Feast (#6604); and in Claude HUD, two display features (#354, #471) and a fix for a crash my #471 caused once Claude Code started sending effort as an object (#491).

• claudemem (Go): local-first memory for coding agents. Markdown files are the record; a SQLite full-text and vector index is a rebuildable cache over them. About 11,000 lines of Go code, not counting blank lines and comments, with nearly as much test code, CI, and release binaries for macOS, Linux, and Windows.
• handoff: a protocol for handing work from a lead model to cheaper models, with resumable spec files and a published small-sample evaluation that includes the run where delegation lost.
• dev-orchestrator: a Claude Code plugin that takes a task from investigation through tests and verification to a pull request.

I turn lessons from production AI systems into smaller, public, inspectable tools. The current thread is reliable agent execution across sessions and teams: durable context, bounded delegation, and workflows that keep evidence close to decisions.

Small demos and hackathon builds, labeled as such in their READMEs: FireSight (a client-side wildfire map on NASA FIRMS feeds), Dipole (a demo agent that deploys a web project to Netlify or Vercel from a chat prompt), and PostPrism (a computer-use prototype whose hosted front end is a simulation).

• Prove before arguing. A small experiment should be able to overturn the plan.
• Fix the bottleneck. Solve the constraint that changes the outcome; defer adjacent cleanup.
• Keep evidence close to the claim. Tests, source, logs, and failure cases beat polished confidence.
• Leave leverage behind. A delivery should make the next run easier to verify, resume, or reuse.

LinkedIn · X

The sidekick answers from this README, the public persona, the six public repositories named above, and the merged pull requests listed above. It replies in a public GitHub issue and does not speak on Zane's behalf.

Open a public question →