My projects span LLM-powered investigations, agent scheduling and cost controls, and tools for AI coding workflows. I write about what I learn along the way: grounding model outputs, designing evals, and testing where systems break.
|
Open-source scheduling and cost controls for AI agents. Controls which work runs, when it runs, and how agents share capacity. |
LLM-powered cloud log investigations, with evaluations for evidence grounding, detection quality, and useful explanations. |
|
A Codex and Claude Code plugin for operational runbooks with verification steps, recovery paths, and execution records. |
A community-built Codex plugin that makes context usage, token consumption, and compaction activity visible. |
- Safe-Agent — Security skills for AI coding agents: skill verification, cost tracking, tool authorization, behavioral anomaly detection, and pre-execution checks.
- Resurface — An iOS app for saving and revisiting links, text, and screenshots.
Golden fixtures, evidence grounding, failure analysis, and regression tests for Flare.
Read on GitHub → · Read on X →
A 20-seed simulation exploring backlog versus fairness—and how queue-depth-only scheduling can starve agents.
Read on GitHub → · Read on X →
Lessons from using two coding agents together: bugs caught, useful disagreements, and workflow limitations.
Read on GitHub → · Read on X →
More writing: Stress-testing Google ADK · Cloud audit-log coverage from inside GitHub
- Tab Review Marker - A free, local-first Chrome extension that adds To Review, Reviewed, and Needs Attention states directly to tab favicons, with notes and AI-ready exports.
- OpenClaw Daily Spending Allowances - Proposed a consumer-friendly, Screen Time-style spending allowance for unattended AI agents, with plugin-owned policy and a privacy-conscious cost-and-provenance contract for OpenClaw core. Awaiting maintainer product review.
- ADK Learning Guide - An interactive deep-onboarding guide to Google's Agent Development Kit, including a LOCO-Agent integration module.
My background spans developer platforms, applied ML, and shipping AI products. At Meta, I led messaging platform products and shipped an LLM co-pilot with answer-quality evaluations. I co-founded PlanetWatchers, helping turn U-Net research into a commercial satellite-imagery platform. Earlier, I built ML-powered observability products at Sumo Logic and AppDynamics.
PlanetWatchers coverage: SpaceNews · Planet.com
- LLM evaluation: grounding, failure analysis, and deciding whether a model or prompt change actually improves the product.
- Agent infrastructure: scheduling, shared capacity, and the trade-offs between cost, latency, and fairness.
- Developer tools: making agent behavior, context usage, and execution easier to inspect and control.
- Research into products: translating model capabilities into useful workflows with measurable acceptance criteria.
- Discovering Anomalies and Root Causes in Applications via Relevant Fields Analysis
- A Message-Passing and Load-Sensitive Contention-Based MAC Protocol for Ad-Hoc Wireless Networks