ProjectBuild an AI Voice Receptionist: the explanation · Module 17: Evaluating a voice agent · part 19 of 23
Part 19 · Module 17: Evaluating a voice agent
Module 17: Evaluating a voice agent
24 min read·
About 40 minutes · Needs: Module 11 (the agent and its tools), Module 9 (Whisper and WER), Module 15 (the safety guards) · Code: backend/evals/run.py, backend/evals/fixtures.py, backend/evals/speech.py, backend/evals/latency.py, backend/evals/voice_common.py
Where this fits
This module (blue) is the Evaluation box: it drives the real agent, Whisper and Kokoro with prepared inputs and checks the results in PostgreSQL.
Tests check that each function does what its author meant. Evaluations check that the whole receptionist behaves well on realistic calls, in three languages, with the model in the loop. This module opens the "Evaluation" box: the datasets, the six suites and the gate that can stop a release.
The rest of this project is yours to keep
This project is bought on its own, once, and stays readable afterwards, including the parts added to it later.