Selected work
A marketplace was growing its top line. The question I framed was whether it was actually healthy — revenue growth hides a broken retention engine very well. Nine normalised tables, so this was answered in SQL rather than in a single dataframe.
- A late order is 7.9× more likely to get 1 star. Orders arriving 15+ days late average 1.72 stars against 4.30 for early deliveries. Late orders are just 6.8% of volume but produce 36.5% of all 1-star reviews.
- I was wrong about the bottleneck — and the data said so. I expected slow seller dispatch. Decomposing delivery into three stages showed last-mile carrier transit takes 9.33 days against 2.80 for seller handover: 74% of the timeline, 3.3× larger. The fix is carrier routing, not seller SLAs.
- A 3.00% repeat purchase rate. 90,557 of 93,358 customers ordered exactly once. The business runs almost entirely on paid acquisition with no retention offset.
- 29 sellers generate 25.5% of GMV. Out of 2,970. Concentration risk invisible on any dashboard reporting totals only.
Does tourism demand track air transport intensity — and can it be forecast reliably? I integrated six real-world datasets covering tourist arrivals and departures, air passenger volumes and aviation safety indicators, then built and compared three forecasting approaches rather than defending a single one.
- Built a reproducible pandas preprocessing pipeline — country-name standardisation across inconsistent sources, forward/backward-fill imputation, duplicate aggregation, IQR-based outlier capping — with the rationale documented for every treatment rather than records dropped silently.
- Engineered year-over-year growth rates, indexed trend series, volatility measures and lag sequences for supervised time-series learning.
- Trained ARIMA, Prophet and an LSTM network, then recommended the most interpretable model for the use case — not simply the lowest error.
- Presented model comparisons to an academic review panel, translating statistical output into plain language for non-specialists.
Smart device usage data analysed end to end through the full six-phase framework (Ask, Prepare, Process, Analyze, Share, Act), closing with marketing recommendations for a wellness technology product team.
- Cleaned and validated the source fitness-tracker dataset in Excel and SQL, resolving formatting inconsistencies and verifying integrity before any analysis ran.
- Surfaced behavioural usage patterns and built the supporting visualisations.
- Delivered a stakeholder presentation with data-backed marketing recommendations.