Take a closer look at what I built

Five specialists, one conversation, one wallet. Why routing by intent wasn't enough on its own, and what happened when the supervisor started calling every number on the screen a lie.

When the knowledge base changes, the rules the agent has to follow get rebuilt on their own, and re-checked before any of them go live.

The agent knows when to stop. The hard part is everything after: who gets pinged, what they see, what the customer sees, and how the agent knows to shut up and not talk over them.

Three products that looked totally unrelated, all sharing one set of core entities. The refactor that stopped a change in one of them from breaking the other two.

You can't check a language model's exact words, but you can check what it must never do. How I turned 'is the agent lying about prices?' into a suite that runs in five minutes and stops a deploy cold.

Every turn, every tool call, every piece of the prompt, saved and searchable. Bug reports stopped being 'the AI is acting weird' and started being a trace id.

No-shows, insurance plans, professionals with different skills, and a calendar that still has to turn a profit. Modeling constraints when the constraints won't sit still.

A crawler, a public barcode database and a whole lot of messy data, so a shop owner can add a product just by pointing their phone at it.

The same traces I use to debug the system also draw it: an isometric office where every figure is a live conversation. How I got 5,000 draw calls down to 8 with instanced meshes.

I fanned out a bunch of finders, then made skeptics try to tear down every claim, and kept whatever was left standing. 41 findings in, 20 out — and why the ones that died matter just as much.