Notes from the build
Working notes on latency, architecture, and the parts that turned out harder than they looked.
Escalation in a prompt is a suggestion, not a rule
Five rules that run before the model — and the tests that keep them quiet on “burn calories”.
A slow AI receptionist can lose a customer before the conversation starts
Cutting response time from 1–2s to ~380ms — and the four changes that did it.
Your AI receptionist shouldn’t wait for Google Calendar, Gmail, and Slack
Queue-backed integrations with retries, and why the database is the source of truth.
Every edge case is a potential bug
20 tests for appointment availability logic, aimed away from the happy path.
The architecture problem isn’t the model
Routing agents, focused toolsets, and what the complex version actually costs.
Good structure is why adding a tool takes three steps
Modular tool registries, and keeping the rest of the codebase ignorant.
Written while building. Each one started as a problem I did not expect to take as long as it did.