PROJECT NOTES / EXPERIMENTS / WHAT WE LEARNED

Deep Dives

Eleven close looks at what we built, what broke, and what each project taught us about building software with agents.

Shape of Done Four months of full-time AI orchestration. The failure mode I see everywhere: agents producing the shape of finished work without doing the work. The Interview A voice and chat experiment that turns a conversation into a structured specification an agent can build from. Boopadoop How a shared AI canvas went from spec to production in one afternoon - with zero human-written application code. Mesopotamian Pottery One rambling question about ancient pottery. One working session. Thirty pages, 13 chronological phases, a live interactive website, and an email-gated PDF. Sidefire The story of building a real-time chat platform with AI agents - the spectacular failure and the lessons that made everything after it possible. Sqzarr A self-hosted GPU transcoder built entirely by agents from a voice interview. Its deficiency report improved the pipeline more than the app improved the server. Fork the Vote Voice in, deployed app out, no human code. The family voting app was the test case. The specification pipeline was the experiment. Johnny Solarseed Educational tools and open-source Home Assistant integrations built around one question: why does solar cost $50K? Codenoscopy A quick experiment with Claude's API turned into a trademarked AI code review tool, built and polished almost entirely through AI coaching. OpenBaseline A thought experiment about making preventive health testing affordable and understandable - born from working inside a clinical diagnostics lab. Smart Home Lab A sprawling Home Assistant installation that doubles as a research platform - where solar optimization, AI-assisted automation, and too many sensors meet.