Language2022Shipped

Customer sentiment model

Composite NLP over call-center transcripts for earlier callback priority.

Problem

Call-center keyword rules were too blunt. Complaint-shaped conversations needed earlier flags so callbacks could be prioritized.

Constraints

  • Transcripts from Nexidia and Genesys with messy real-world audio ASR
  • Had to beat the prior keyword-only rules system, not just look clever
  • Output had to fit existing callback workflows

What shipped

A composite NLP stack — topic modeling, sentiment scoring, and keyword extraction — with a TensorFlow LSTM head, producing a single priority signal for ops.

Outcome

Complaint-shaped conversations surfaced earlier than keyword rules alone, feeding the same callback queue the floor already used.

More cases

All works →