Contact Intent & Complaints MLAXA UK

Forecasting, classification and summarisation models for AXA's contact-centre and complaints teams. Improving the digital self-service journey cut calls by 30–60%.

In numbers

30–60%: reduction in calls
£0.4–0.8m: estimated business benefit
813 → 527: call-volume forecast RMSE

What I did

1.
Worked with Accenture on contact-intent and channel-switch optimisation, enriching unstructured raw data from GCP for ML modelling; EDA and automated reporting improved the digital self-service journey, cutting calls by 30–60% (est. £0.4–0.8m benefit).
2.
Built a custom LSTM and fine-tuned a Facebook Prophet model to forecast call-centre volumes, improving RMSE from 813.35 to 527.1.
3.
Developed and deployed a TensorFlow deep neural network on PySpark to classify complaint outcomes (upheld / not upheld / partially upheld), with Optuna tuning and MLflow experiment tracking introduced across the team.
4.
Benchmarked BERT and T5-small for complaint summarisation and fine-tuned T5-small for production, significantly reducing agent review time.
5.
Implemented a random-forest regression model to predict compensation amounts, delivering 15–20% cost savings.

Notes

A run of connected projects across AXA UK's customer-operations data: forecasting how many people will call, understanding why they call, and helping agents resolve complaints faster.

I led daily stand-ups and sprint planning for a small cross-team squad and delivered a Plotly dashboard that surfaced the complaint-classification and summarisation outputs to stakeholders.

Built with

Python · GCP · PySpark · TensorFlow · LSTM · Prophet · Hugging Face · T5 · Optuna · MLflow · Plotly

Next project

Trading Platform & GCP Data Pipeline →