I build open-source tooling for AI agent observability — making the cost and behavior of LLM agents measurable, honest, and enforceable in CI.
⚡ Wattage — find the tokens your AI agent wasted, and fail the PR when it gets more expensive
Reads the data your agents already produce — Claude Code sessions and OpenTelemetry GenAI traces — prices every call against a verified, dated 52-model pricing snapshot, runs ten waste-pattern detectors (uncached prefixes, thrashing loops, retry storms, oversized tool results…), and ships the thing no dashboard has: a CI cost-regression gate that fails the build when an agent quietly gets more expensive.
uvx wattage demo # findings-rich report in 30 seconds, zero setup
uvx wattage report --claude-code # your latest Claude Code sessionThe design principle throughout: never fabricate a number. Unpriced models fail loudly instead of being guessed, heuristic estimates are labeled and can never fail a build, and every benchmark in the README reproduces from the shipped code.
- 🖥️ computer-use-automation — an LLM discovers a flow on a legacy UI once; deterministic code compiles that run into a typed, content-hashed capability that replays with no model in the loop — behind one policy gate, with a human approving anything irreversible.
- 🧠 STAR+FAR — continual learning for LLMs: sparse temporal LoRA-adapter routing with freshness-aware replay, for staying fresh without forgetting under a compute budget.
- 🌪️ Crisis-Management Relief Coordinator — a multi-agent NLP + CV system fusing radar data, social-media signals, and FEMA guidance for real-time tornado response.
- 🛰️ Nighttime-Satellite-Imagery Economic Predictor — predicting GDP growth from VIIRS nighttime-light intensity correlated with World Bank data.
- Catching the AI Agent Failure Mode Your Loop Guard Can't See — why exact-match loop guards structurally miss fuzzy retries, oscillation, and productive-looking stalls, and how Wattage's convergence engine measures progress instead.
AI agents · LLM cost & observability · OpenTelemetry · continual learning · applied ML