🐙 Agents that use computers | 🔬 LLM interpretability | 🚀 Building Openmart AveraLabs

I build agents that touch the real world

• 🐙 Octopus – Drives any Linux app through screenshots and kernel-level input events. No Playwright, no CDP, no accessibility tree — just pixels and a mouse. Emits /dev/uinput events byte-for-byte identical to a USB mouse, so nothing downstream can tell it isn't a hand on a desk.
• 🤖 Chrome-GPT – LangChain + Selenium agent that scrolls, clicks, and fills forms across a whole Chrome session. Zero-shot, BabyAGI, and AutoGPT agent types.

• 📱 OpenCapy – Go daemon that mirrors your Claude Code sessions to your phone. Approve prompts, browse diffs, drop into a live PTY — over relay, Tailscale, or SSH.
• 🍎 opencapy-ios – The SwiftUI client. Interactive terminal, native image paste, voice input, Live Activities.
• 🍺 homebrew-opencapy – brew tap richardyc/opencapy

• 🔬 llm-backdoor-report – Systematic backdoor hunting in four fine-tuned models with hidden triggers. SVD weight-diff analysis found the modified matrices; logit lens found the real mechanism — not an if/then injection, but a graded MLP suppression failure at layer 60. Ablate it and the hidden token jumps from 18% to 83%.

• 🏠 Jarvis Home – Always-listening speech-to-speech assistant running as a custom APK on a second-hand Lenovo smart display whose vendor backend died years ago. On-device wake word, WebRTC AEC3 at ~27 dB for real barge-in, and I2C init tables replayed by hand because the debug firmware never brings the ADC or amp up. 40-second demo →

• 🎙️ AveraLabs – Building the world's most realistic full-duplex voice model. no turn-taking, no VAD gating, no waiting for you to stop talking. Real interruptions, backchannels, and overlap, the way people actually talk.

• 🌐 openmart.com
• 🐦 @RealRichomie