Patter SDK Guide to Building a Restaurant Booking Phone Agent with Dynamic Variables, Guardrails, Latency Dashboards, and Eval Checks

There's a specific kind of tutorial that voice-AI builders have been missing: one that doesn't require a Twilio account, a live phone number, or actual API credits just to see whether an agent's logic holds up. MarkTechPost's walkthrough of the Patter SDK fills that gap by building a restaurant booking assistant that runs entirely in-process, complete with fake speech-to-text, fake text-to-speech, and a scripted call flow you can rerun as many times as you want without spending a cent.

The setup is refreshingly concrete. The tutorial defines caller variables like customer name and loyalty tier, then registers five callable tools — check_availability, book_table, get_hours, lookup_reservation, and transfer_to_human — against an in-memory backend seeded with specific slot counts (six open tables tonight, zero on Friday evening) and one existing reservation under confirmation code AC8842. Every simulated call resets that backend from scratch, which is the kind of detail that matters if you actually care about reproducible tests rather than demo-day vibes.

What stands out is how much attention goes to guardrails before any conversation logic gets written. The pipeline redacts emails and phone numbers with regex, swaps internal customer IDs for generic phrasing, strips mild profanity, blocks off-topic requests like medical or legal questions with a canned redirect, and trims replies down to two sentences max so the agent doesn't ramble on a phone call. Those five guardrail functions run in sequence on every output, which is a sane default for anything that's going to talk to real humans on a real line eventually.

The agent brain itself is deliberately deterministic rather than an LLM free-for-all — it's a chain of regex checks and state tracking that walks a caller through booking a table one field at a time: party size, then date, then seating slot, then name, then a tool call to actually reserve it. There's an optional hook to defer small talk to a real OpenAI model if you flip USE_REAL_LLM and have a key set, but the core flow doesn't need one, which makes the whole thing testable without nondeterminism creeping in. The tutorial also tracks latency at each stage — STT, LLM, tool execution, TTS — per turn, plus cost, setting up for the regression-style eval checks mentioned at the end that validate the system stays behaviorally consistent across runs.

The getpatter package itself is young, per the tutorial's own admission — young enough that the code checks the installed API at runtime rather than assuming a fixed signature, which tells you this is very much a moving target still finding its shape.