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 agents: memory, delegation, and evidence-driven development.

Explore the work ↓ · LinkedIn · GitHub

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.

Current public proof: claudemem preserves context across coding-agent sessions, handoff measures when delegated execution helps and when coordination cost outweighs the benefit, and dev-orchestrator keeps investigation, tests, verification, and shipping in one resumable workflow.

The product experiments below apply the same standard to computer use, real-time data, and deployment without presenting prototypes as production systems.

• claudemem — persistent memory for coding agents, using portable Markdown records plus searchable indexing across sessions.
• handoff — a spec-and-ledger protocol for token-tiered delegation, published with its pre-registered evaluation and failure direction.
• dev-orchestrator — an end-to-end development workflow that connects investigation, planning, tests, verification, shipping, hooks, and file-backed state.

• PostPrism — a hackathon prototype with a front-end simulation and an experimental backend for parallel computer-use agents.
• FireSight — a client-side wildfire map built around NASA FIRMS feeds and Leaflet.
• Dipole — a conversational deployment assistant for Netlify and Vercel with streamed progress and diagnostics.

Upstream work in other people's repositories:

• jarrodwatts/claude-hud — three merged PRs in the 26k-star Claude Code HUD: configurable model display (#354), effort-level display in the model bracket (#471), and a schema-compatibility fix for newer Claude Code releases (#491).
• modelcontextprotocol/typescript-sdk — a reproducible bug report on v2 declaration source maps (#2491), built from published-tarball forensics.
• letta-ai/letta — three tested fixes staged on a public fork and linked from issues #3310, #3399, and #3390: provider rate-limit attribution, slash-label routing with route-shadow regression tests, and a missing client timeout.
• bhimamalbhage/lightup — merged multi-chain agent design work (#3, #4).
• nextbound/bragi-canvas — an Obsidian canvas plugin; merged upstream bugfix (#26).

• 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.

The canonical profile stays text-first. Three complete visual interpretations live in the profile design gallery: Console, Constellation, and Field Notes. They are design studies, not alternate claims.

LinkedIn · GitHub · X