Los Angeles. I build open-source AI tools, and I make AIs argue with each other until one of them ships something.
- Open Council: three AIs argue over your spec until they agree, then one ships it as a tested, gated PR. It never says done when it isn't. Watch a real council →
- OmniGlass: AI that acts on your screen, not just reads it. Snip anything, get buttons that execute. MCP plugins, sandboxed, local LLM.
- Anie: reads a news article and flags manipulation, bias and psyops. Live at goanie.com.
- OpenCouncil: Lead PM, Chief Engineer and Chief Architect seats deliberate a PRD to unanimous sign-off; you approve; it opens the PR. Claude Code plugin, CLI and agent skill.
- council-playground: the small repo I point councils at to test solo mode end to end.
- OmniGlass: Tauri and Rust desktop app. Screen pixels in, MCP tool calls out.
- NARS-JARVIS: a local cognitive assistant that pairs a NARS/ONA symbolic brain with a local LLM for language.
- Anie: forensic analysis of news articles. FastAPI backend, Chrome extension, mobile app.
- project-eagle-eye: real-time WiFi CSI sensing that renders people as 3D skeletons on a floor plan.
- Running Open Council on real repos and publishing every run, stalls included
- Making multi-agent coding boring enough to trust: gates, sign-offs, no false DONEs
- Writing about what breaks at ashkan.us
Plausible isn't the same as approvable.
Things my councils taught me
- In ten runs in one day, every council pointed at my test repo caught that
npm testhad been running zero tests on Node 22. Nobody asked them to look. - A PR can pass every check and still not build against main. Now the harness syncs with the base before it verifies.
- The most useful bug report is a council that went wrong.



