Autonomous agents, local LLM infrastructure, research tools. Built to work. Not to demo well and break on real tasks. Every project here is something I needed and nobody had built.
Every autonomous web agent I tried worked on demo sites and collapsed on anything real: Cloudflare, dynamic JS, anti-bot measures, login walls. So I built one that doesn't.
ReAct loop. A DOM walker that extracts semantic structure the LLM can actually reason about instead of drowning in raw HTML. Stealth Playwright. 2,904 tests because autonomous agents fail silently and I needed to trust it.
Runtime-agnostic communication mesh for MCP clients. 21 tools, shared memory with FTS5 search, broadcast channels, task delegation, stable identities, presence tracking, and cross-machine coordination over Tailscale with bearer auth.
View project →Neuroscience-grounded emotional architecture for AI agents. 6-phase pipeline: PAD dimensions, Panksepp drives, somatic markers, emotion regulation, mirror neuron contagion, Barrett granularity. Persistent deep reflection daemon thinks between conversations. Generates behavioral directives from its own reflections.
View project →1.3M file cross-cultural mythology and history corpus across 69 traditions. Full RAG pipeline with LlamaIndex and ChromaDB. Automated ingestion via Ghost Hand (autonomous browser agent). Built as the training substrate for local LLM fine-tuning.
View project →GitHub activity dashboard for your terminal. Contribution heatmap, starred repos, live notifications — all in a BubbleTea TUI. No browser required. Configurable refresh, token auth, clean column layout.
Tool infrastructure for Claude Code. YouTube intelligence (7 tools, Whisper fallback, frame extraction), DevMail (IMAP/SMTP email access), Grayson (job outreach pipeline with email verification), Audio analysis (19 tools, CLAP, Demucs, Whisper), crypto signals, and more.
View project →Self-improving uncensored reasoning agent. Hermes 3 405B via OpenRouter. Connected to the claude-peers network. After every response, runs self-critique, scores its own reasoning quality, and generates lessons that compound into future conversations. 15 lessons learned in first training session.
View project →Autonomous agent loop engine. Runs Claude Code sessions iteratively against real completion criteria: tests, file checks, grep assertions. Runs until done or escalates. Solves the "always in beta" problem.
View project →Agent evaluation framework. Define what your agent should do, run it against frozen baselines, catch regressions before they hit production. Built because existing tools test chat quality, not whether your agent still calculates refunds correctly after a model swap.
Research project exploring whether small, purpose-trained models can match large general models on structured prediction tasks. Working with historical texts to study how political systems change over time.
View project →Autonomous trading agent running on Base chain. Cycles on a schedule, scores market signals, makes decisions without human intervention. Paper trading for safety — proof of architecture before real capital.
The missing half of Claude Code's voice mode. Claude talks back with Kokoro TTS and real-time karaoke-style word highlighting in your terminal. Fully local. Zero API keys. One file.
I build autonomous agents, LLM pipelines, multi-agent systems, and research infrastructure. If you're working on something that needs to coordinate AI agents, handle adversarial web environments, or run without cloud dependency. That's the area.