10 years in production systems at scale. Designed SLO platforms serving 1,000+ microservices, built topology-aware anomaly detection and 5-minute automated RCA across billions in daily transaction volume. Led a company-wide JVM migration (800+ services, ZingJDK to OpenJDK 21 ZGC, GC pause from tens-of-ms to <1ms).

Now I split my time between reliability architecture and building AI agent tooling. I believe the barrier to building has dissolved — what matters is the restlessness to keep tinkering, treating every new tool as a puzzle to solve before the day is over.

Beyond open source, I also build in private: an AI-coordinated quant trading system (Rust engine, 4 autonomous AI departments, Sharpe 4.9, 12.8M data points), an options portfolio manager (FastAPI, strategy scanner, E2E tested), and a physics-based golf simulator (RK4 integration, USGA Stimpmeter calibrated, live on Vercel). All shipped with AI as a force multiplier.

skillpm Universal package manager for AI agent skills. One CLI to install, sync, and secure skills across 10+ agent runtimes — Claude Code, Codex, Gemini CLI, Copilot, Cursor, and more.

lindy-orchestrator AI agent orchestration grounded in harness engineering — DAG scheduling, QA gates, worktree isolation, multi-provider dispatch, git-native coordination.

symphony-claude Turns project work into isolated, autonomous implementation runs. Teams manage work instead of supervising coding agents.

microsoft/VibeVoice · contributor Fixed a tensor shape mismatch in the streaming TTS inference engine that broke SDPA attention on MPS/CPU (#303). Root-caused to a missing query_length in the KV cache mask contract for transformers 4.57.