Lab

Rough cuts and experiments. Artifacts from exploring various domains.

JorgeRefinementA self-hosted, sandboxed Hermes personal agent that tames my messy smart-home setup — and runs my calendar, planning, and knowledge base, even ordering the groceries. Grumpy by design.hermeshome assistantopenrouterqwenminimaxelevenlabs scribegemini ttstelegramtailscaledockerhetznergitops

Bullhorn↗Post to X from inside Claude Code — open-source, BYOK, write-only plugin. Draft from your session, review, approve, post.Claude CodeX APIMCP

Polymarket Trading BotRefinementAutonomous prediction-market trading — sports copy-trading plus weather arbitrage.polymarketnode.jstypescriptautomatic trading

Deepgram Eval HarnessRefinementEval harness for a production Deepgram voice agent — intent accuracy and end-to-end latency.deepgramvoiceagentevalharnessmulti-turn

Prompt Injection Demo↗Interactive demo of a prompt-injection attack: instructions hidden in markdown input override the system prompt and exfiltrate context data through image requests.minimalprompt injectionagentsecurity

Plus a back-catalogue of pre-2025 experiments — unpublished.