CALL-QA

THE PROBLEM

A head of sales can’t listen to every call, and a model that “scores” calls without evidence is just one more opinion. On top of that, customers read card numbers out loud — and those must never end up in an AI provider’s logs.

WHAT I BUILT

• A pipeline from speech-to-text output to a scored call: roles, masking, scoring, checks — with three attempts per stage and an alert when the provider keeps failing.
• Speaker roles by channel on stereo recordings and by phrases on mono.
• A seven-criterion scorecard with weights. The model scores each criterion; the server computes the total, and the model never sees the weights.
• A team view: the manager × criterion heat map, red flags and a coaching focus per manager.

HOW IT’S CHECKED

• Card numbers, phones and emails are masked before anything is stored or sent to the model — including digits read out as words, in English and Russian.
• Every score cites utterance ids. Ids that aren’t in the call are dropped, and a score that credits the manager has to quote the manager’s own words — otherwise it doesn’t count.
• Hard rules are code: a discount over 5% without approval, a guaranteed delivery date, card details on the phone. A rule hit forces compliance to 0, whatever the model says.
• The evidence is on the page: each line of the transcript carries the scores that cite it, and a mark jumps to its scorecard row.

WHAT THE DEMO SHOWS

Eleven scripted calls for the same fictional Northwind Air and four managers. Speech-to-text is simulated. Nine calls are scored; two are skipped without a model call, a voicemail and a seven-second call.