January 2023 – June 2023
Setel Ventures was my first real internship. January to June 2023, based in Kuala Lumpur. Fresh out of my first semester at SIT, no production experience, and suddenly responsible for ML models that affected real business outcomes.
It was a sharp start. That was the point.
Setel is Malaysia's leading fuel and mobility super-app — initially launched by PETRONAS. Millions of users, petrol pump integrations, gig-worker partnerships, loyalty programmes. The data is large, messy, and consequential.
The first project: rebuild the Customer Topup Segmentation Model. The existing model was version 1 — it worked, but it wasn't capturing recent behavioural shifts in how customers topped up.
I rebuilt it from scratch (version 2) with updated features and multiple rounds of K-Means clustering to get sharper segment boundaries. Result: increased top-up business value by 30%.
The follow-on: a Customer Fuel Segmentation Model using the same approach but applied to fuel purchasing behaviour. More detailed class definitions, more rounds of clustering iteration. Increased fuel business value by 30%.
Setel had a gig-worker programme — delivery riders, drivers, freelancers. The challenge: limited labelled data. Most users weren't explicitly identified as gig workers, but their behavioural patterns were distinctive.
I built a semi-supervised learning model to identify likely gig workers from unlabelled data, using the small labelled set as anchors. 80% prediction accuracy on the holdout set. The output fed directly into targeted campaign decisions.
The fraud detection pipeline was functional but slow — inefficient code that had accumulated over time. I refactored the core processing logic, eliminating redundant operations and cleaning up the architecture. Script-running duration reduced by 20%.
| Tool | Use |
|---|---|
| Python | All modelling work |
| Pandas | Data manipulation |
| Scikit-Learn | K-Means, semi-supervised learning |
| Data Visualization | Matplotlib, Seaborn |
Three things that stuck:
Numbers matter. Every model change has a business outcome attached. 30% uplift in top-up value isn't an abstract metric — it's revenue. Learning to connect model decisions to business outcomes early made everything that came after clearer.
Semi-supervised learning is underrated. Most real-world datasets are partially labelled at best. The gig-worker model was my first serious exposure to working with limited labels — and it showed me that the constraint of "not enough data" is often solvable with the right framing.
Refactoring is real work. The fraud detection cleanup wasn't glamorous, but reducing script time by 20% across a production pipeline is tangible. Clean code isn't just aesthetics — it has operational value.
Setel was where I learned that data science is engineering, not just analysis. That framing carried through everything else.