What I’ve learned building and deploying AI systems across four companies.
The agent isn’t the product
The most common mistake I see: teams spend months building the perfect agent, then realize they have no way to deploy it safely. The agent is 20% of the work. The other 80% is guardrails, monitoring, fallbacks, and observability.
Plan first, code second
Every successful AI feature I’ve shipped started with a written plan — not a PRD, not a spec, just a clear description of what the system should do and what happens when it fails.
Production is the only test environment that matters
You can test in staging all you want. Real users will find edge cases you never imagined. Ship behind a feature flag, watch the metrics, iterate fast.
Open source as leverage
My most impactful projects — even_glasses (79★), telegramGPT (52★), G1 Voice AI (25★) — started as tools I needed myself. Open source forces you to write clean APIs, good docs, and handle edge cases. It’s the best forcing function for quality.
What’s next
Building AI agents that operate safely in production — with guardrails, snapshots, and human-in-the-loop review. If that sounds interesting, let’s talk.