Things built outside work — all open on GitHub. The market-analysis system behind the research window is the fourth.
AI-agent observability platform — a “Datadog for AI agents.”
Custom OpenTelemetry collector pipeline (span classification, real-time cost enforcement, DAG topology inference) feeding ClickHouse for traces and Postgres for graph + alerting. Python and TypeScript SDKs, a Next.js dashboard with topology visualization and session tracking, LLM-as-judge eval scoring, two-tier agent loop detection, and multi-signal alerting.
A suite of 10 MCP servers (41 tools) for options-flow analysis.
Tools for analyzing options flow, dark-pool activity, open-interest changes, and market signals from Unusual Whales data exports — signal-confluence scoring, market-regime detection (SPY/VIX trend analysis), portfolio correlation analysis, and automated options-strategy suggestions with Yahoo Finance integration.
Open-source MCP server that recommends the right Claude model per prompt.
Dynamically classifies prompts using claude-haiku-4-5 (~100 ms, ~$0.0001/call) and recommends the optimal Claude model by task type and complexity. Contributed a plugin to the official Anthropic claude-code repository; cited as a working reference implementation for hook-based model routing.