AI Call Quality Control System
An automated pipeline that listens to every customer call and tells a moving company whether the quote can be trusted — no human listens to a single recording.
Project Overview
A small team of agents handled inbound moving enquiries by phone, then sent quotes manually a few hours later. The quote and the call sometimes disagreed — an agent mishears a floor number, packing materials discussed on the call vanish from the quote — with no review process and no coaching data to fix it.
We built a system that transcribes and analyses every call automatically, connecting Ringover, RunPod (WhisperX), and Gemini 2.5 Pro into one pipeline with no manual trigger.
How It Works
A Bull job polls Ringover every 15 minutes for new completed calls and matches them to cases by normalised phone number. Matched recordings are transcribed with WhisperX on RunPod, producing a labelled, speaker-diarized transcript. Gemini 2.5 Pro then compares the transcript against the quote using a versioned prompt and returns a structured verdict — YES, NO, ESTIMATED_QUOTE, or MULTIPLE_QUOTES — with a written explanation and agent coaching notes.
- Poll & match calls to cases every 15 minutes
- WhisperX transcription with speaker diarization on RunPod
- Gemini 2.5 Pro structured verdict + agent coaching
- Redis + Bull async queue with retry on failure
Screens
Outcome
The system now processes new calls continuously with a roughly 15-minute call-to-verdict latency. Of 122 fully analysed cases, 47 came back NO — real discrepancies that previously had zero detection mechanism. Agent coaching now has real data behind it for the first time.
- 47 real discrepancies surfaced out of 122 analysed cases
- Agent performance scoring with real call data
- 35 estimated-quote cases flagged for follow-up
- Runs unattended — PM2 + scheduled polling, no manual triggers
Tech Highlights
Overview
Need every call checked automatically, not spot-checked by hand?
Let's discuss the best stack for your business