Skip to content
← Back to work
NLPClassical ML

NLP Customer Intelligence (Vodafone)

Data Scientist · 2022-2023

Mining Vodafone's ignored open-text customer feedback for insight nobody else would touch, and what fighting a big org's bureaucracy taught me about where I do my best work.

Setup

My time at Vodafone was one of the pivotal points in my career: the place where I first really felt the weight of hierarchy. We handled millions of records a day, query after query to optimize feature extraction and evaluation for the models in production: churn detection, recommenders, NPS predictors, and more.

But one data source sat untouched: the Verbatim. These were the open-ended comments from every customer the NPS team called about their experience, free text written alongside the 1-to-10 scores for the quantitative questions. Having worked on something similar during my time at Bocconi, I knew open text like this was a goldmine. Nobody was mining it.

Why it was hard

The problem was never the modeling. It was the organization. I didn't want to spend months hunting for features in an ocean of data I would never fully master, all for the faint hope of moving a model's accuracy by 0.0001%. I wanted to do something that mattered. So I made extracting value from the Verbatim my mission, and that is where the fight began.

In a large organization you cannot just try things. As a junior, every step meant going through the nine circles of Dante. I couldn't access Git or pull open-source models. Every day I opened a ticket so the security team would let me use basic Python libraries like NLTK or spaCy. GPT models? Out of the question, too risky and unconventional (this was before ChatGPT, though GPT-3 already existed). I had to fight marketing for the NPS data, fight data engineering to make the ETL for open text efficient, and fight the cloud architect for a container image powerful enough to run language models. A junior, three months in, causing trouble for everyone.

Impact

The work changed how the organization treated NPS: the quantitative side got real investment, a dedicated model, and a team around it. I had shown, with evidence, that customer feedback held value well beyond a single score.

Why it didn't work

But the open text itself, the thing I had fought for, went nowhere. Even with my manager genuinely in my corner, the chain stalled: he convinced his manager, who convinced his manager, and somewhere near the top it died. After six months the answer came back: not interested in open text for now. A hierarchy where doing your assigned job counted for more than doing the job that actually mattered.

I learned an enormous amount at Vodafone. The most important lesson was about fit: I do my best work where curiosity sets the agenda, not the org chart. When there is no curiosity, there is no life.

Takeaways