Juan Carlos Jiménez
← All work

03 / 11 · SURVEYS

Survey classifier

A Spanish BERT model, fine-tuned to classify customer surveys and to know when to trust itself.

Context
Telefónica / Movistar Plus
Role
Modelling and deployment
Year
2025
Stack
BETO · Fine-tuning · scikit-learn · Python

The problem

A steady stream of free-text surveys, three imbalanced classes, and a team reading them by hand.

What I did

I fine-tuned BETO, ensembled it with a TF-IDF model, and added confidence-based routing: clear cases resolve on their own, doubtful ones go to a person.

What changed

Most surveys now resolve without anyone reading them, and people only see the cases where their judgement matters.