{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/5750342a89ab4cafb318a54be992937b\" frameborder=\"0\" width=\"1670\" height=\"1252\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":1252,"width":1670,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":1252,"thumbnail_width":1670,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/5750342a89ab4cafb318a54be992937b-d35f0b8e4179df9d.gif","duration":622.672,"title":"Pretty Good AI VoicePod Bug Detection Walkthrough","description":"This Loom presents an automated patient caller that uses two phone-connected AI pods to converse and then generates bug reports after each call. The system streams Twilio call audio through a small API server directly into OpenAI’s Realtime API for low latency, avoiding a multi-step transcribe think speak pipeline. The emphasized portion is the bug detection pipeline in analyzes.py, which drafts bug reports by sampling three independent passes against a fixed rubric and merging results to reduce missed bugs, then grounds findings in lightweight ground-truth clinic facts such as hours, addresses, and provider names. An example high-severity bug is a contradiction where the agent first suggests an appointment is booked and later claims there are no open cases, and the UI lets viewers play audio, view transcripts, and see the bug report per scenario. "}