Lessons from building production AI systems — what worked, what failed, and what I'd do differently.
Getting to 72% was the easy part. The 28% that still needs human handling — ambiguous intent, emotional escalation, multi-step state — is where the real engineering lives.
Groq Llama 3.3 for speed, Claude for judgment, Ollama for free deterministic tasks. Routing correctly across them is the optimization — not picking the smartest model.
The protocol is the easy part. Idempotency, rollback, and audit trails — those are what separate a demo agent from one you can trust with internal operations.