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

A pipeline designed to pull batch audio from cloud storage, transcribe it with Whisper, and score each transcript with a multi-label toxicity classifier (toxic-bert) across six categories — toxic, severe, obscene, threat, insult, identity hate — flagging calls above a threshold, with their transcript, for QA review. The design covers the data model, storage architecture, and the compute budget for the full archive; the download, transcription, and scoring functions are built.

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

Signature records from roughly 129,000 timecard sign-offs are extracted, scoped to a three-month window, and catalogued per signer, with signer identities normalized from free-text names so each signer’s history can be assembled. That catalogue is the training set for the next step: a convolutional neural network learned per signer, so a signature that doesn’t match its claimed signer can be flagged for review.