Every sales call scored against the checklist

The problem

The sales team of a new-build real estate agency in Kazakhstan was on the phone all day. The head of sales could listen to only a small sample of calls, so most script mistakes and badly handled objections went unnoticed. The agency needed every call checked against the team's QA checklist, without the head of sales spending hours on audio.

What I built

A Python service backed by PostgreSQL. Each call recording goes through two stages:

1. Transcription with Yandex SpeechKit, which turns the recording into text.
2. Scoring: a language model called through the OpenAI API reads the transcript and scores the call against that checklist.

Recordings move through a queue and are handled by background workers, with retries and logging. A recording that fails is visible and gets processed again, so no call quietly drops out of the review.

The result

The pipeline ran in production on the agency's calls. Review stopped being a sample: every call got a score, not only the few the head of sales had time to hear.

Scores replaced a large share of manual call review. The head of sales read the results instead of listening to hours of recordings.

For the same agency I built an outbound AI voice agent. Packages and prices for such systems are on the call and document analysis page.