Hands on the keyboard

Problem

Recorded customer-service calls hold the earliest signal of a relationship going wrong, but nobody can listen to tens of thousands of them. The question was whether a pipeline could surface the handful that need a human.

Approach

Batch audio is pulled from a folder or bucket, transcribed with Whisper, and each transcript is scored by a multi-label toxicity classifier (toxic-bert) across six categories — toxic, severe, obscene, threat, insult, identity hate. Any label above a threshold flags the call, transcript attached, for QA review. The public version runs end to end on synthetic calls: eight scripted calls, two hostile, all eight triaged correctly — and the write-up reports the first draft’s miss (polite agent turns diluting the score) and what it implies for real archives.

Problem

A nurse’s timecard is signed off by facility staff. A signature that doesn’t belong to the person it claims to is a billing-integrity problem that is nearly impossible to catch by eye at volume.

Approach

First a signer catalogue: every signature attributed to a normalized signer identity so each signer’s history can be assembled. Then a verifier, completed on the public CEDAR dataset: a Siamese CNN that embeds a questioned signature and a signer’s known signature and flags the pair when they sit too far apart. The dataset’s famous leak — genuine and forged scans differ in background texture, enough for a raw model to score 100% by recognizing the scanner — was measured with a questioned-image-only diagnostic and neutralized before training; results are five-fold writer cross-validated rather than a single lucky split.