A selection of my strongest data analytics, business intelligence, and machine learning work.
A browser-based, no-code AutoML tool that takes any CSV through a full end-to-end pipeline: data intake, automated cleaning, model training, and live prediction, all running client-side. Includes confusion matrices, feature importance charts, and an integrated co-pilot assistant.
Explored workforce data through an HR equity lens. Built entirely in Excel, importing and cleaning raw HR records, using pivot tables and conditional logic to surface patterns in headcount, department distribution, and demographic breakdowns, delivered via an interactive dashboard.
Supervised ML classification using the Titanic dataset. Pipeline covers EDA, feature engineering (FamilySize, IsAlone, binned age/fare), missing value handling, and categorical encoding. Achieved 80.4% accuracy with Logistic Regression. Gender was the strongest predictor; class acted as a proxy for resource access.
End-to-end analysis of a global movie ratings dataset across genre, country, age, and subscription type — built in Excel. Key metrics: $1.34M revenue vs $482K budget, 4.20/5 rating. Interactive dashboard with slicers for genre, country, age, gender, subscription type, and release year.
The same dataset reimplemented in Power BI with deeper technical depth. Custom DAX measures dynamically calculate KPI text-based winners. Power Query handles schema optimisation, type casting, null handling, and a lean data model. Drama is top-ROI genre; France generates maximum revenue.
Production-grade 3-page Power BI dashboard reverse-engineering the Netflix global catalog. Rigorous Power Query ETL: multi-value fields (country, director, genre) split into distinct rows; duration and date fields engineered for accurate metric isolation. Pages cover content strategy, talent networks, and upload seasonality.
A production-grade 5-page Power BI dashboard analysing stockout patterns, demand behaviour, and stocking efficiency across 5 Chicken Republic branches in Lagos (Ajah, Lekki, Ikeja, Surulere, and Yaba). The dashboard was built to give operations and supply chain teams actionable visibility into where, when, and why stockouts occur and what it costs the business.
Key findings: an overall stockout rate of 0.22, ₦12M in total lost revenue, and 401 stockout days across the dataset. Ajah is the worst-performing branch (0.29), Fried Chicken (Full) is the most frequently stocked-out item, and stockouts peak between hours 19–20. Saturday drives the highest sales volume, and promotions marginally reduce stockout rates (0.21 vs 0.22 without promo). Suggested opening stock of 119.59 units with a stock adjustment of -6.28 per cycle.
The 5 pages cover: Executive Stockout Summary, Stockout Overview, Demand Patterns, External Drivers (weather and promotions), and Stocking Efficiency, each with interactive slicers for branch, item, and date.
An end-to-end Power BI dashboard transforming complex IoT sensor logs from a commercial HVAC chiller plant into actionable energy insights, helping facility managers identify inefficiencies, reduce operational costs, and optimise building cooling performance. Built on a star schema data model with custom DAX measures and a dark-mode visual layout.
Key metrics: 0.48 kW/RT average efficiency, 1.73M kWh total energy consumption, 520.94 RT average building load, and a 10% inefficient rate across 14K logged intervals with anomaly spikes reaching 1.51 kW/RT. Building demand peaks sharply at 08:00 AM and ramps down after 08:00 PM, exposing clear off-hour setback opportunities.
The 4 pages cover: Executive Summary (baseline KPIs and multi-year trend analysis), Weather Impact (efficiency vs. outdoor temperature and humidity), Time Pattern Analysis (hourly, daily, and weekday/weekend load behaviour), and Energy Efficiency & Anomalies (flagging 1K inefficient intervals and isolating system performance spikes).
An end-to-end predictive maintenance system bridging mechanical engineering domain knowledge with machine learning, simulating a physics-informed sensor fleet of 19 industrial units and building a failure-prediction pipeline that flags at-risk equipment 7 days before failure. Results feed a 4-page Power BI dashboard designed for fleet managers and maintenance planners.
The synthetic dataset isn't random noise, it's built on realistic degradation physics: vibration and temperature climbing on an accelerating curve in the weeks before failure, current draw rising with mechanical resistance, pump pressure dropping under cavitation. 41,363 sensor readings across 4 equipment types, 8 failure modes, and an 18-month simulation window. A Random Forest Classifier with time-aware train/test split (no data leakage) achieves ROC-AUC 0.9985, 97% recall, and catches 799 of 825 pre-failure windows in the test set. Top features: current vibration level, 3-day vibration slope, and rolling vibration volatility consistent with real-world condition monitoring literature.
The Power BI dashboard covers: Fleet Overview (KPIs, risk distribution, top-5 at-risk units), Equipment Deep Dive (live sensor trends with danger-zone lines per unit), Failure Risk & Predictions (fleet risk ranking matrix and failure mode breakdown), and Maintenance Impact (MTBF and cost avoidance from predictive vs. reactive strategy).
An end-to-end churn prediction system for a simulated digital payments / agent-banking platform, modeled on the transaction patterns of Nigerian fintechs like Moniepoint. Applied RFM (Recency, Frequency, Monetary) and behavioral trend analysis to flag at-risk users 30 days before they churn, backed by a 4-page Power BI dashboard for fleet-level and user-level decision-making. Built on a fully synthetic dataset with no affiliation to Moniepoint, modeled on the industry it operates in.
46,676 monthly user-behavior records across 6,000 users over an 18-month window, spanning 3 KYC tiers, 2 channels (App and USSD), and 3 engagement segments, with a ~54% eventual churn rate. Users decline through a realistic 2–4 month behavioral ramp rather than a random flag: transaction frequency and value drop, failed-transaction rate climbs, and wallet balance drifts toward zero. A Random Forest Classifier with a time-aware train/test split (no data leakage) achieves ROC-AUC 0.9990, 100% recall, and 89% precision on the churn class, catching all 1,739 pre-churn cases in the test set.
The top predictive signal was failed transaction rate and its 3-month trend, ahead of simple activity decline, a business insight suggesting retention efforts should prioritize fixing transaction failures over generic re-engagement campaigns. The Power BI dashboard covers Overview (KPIs, churn trend, risk distribution), Churn Risk Ranking (user-level risk matrix by KYC tier and engagement), Behavioral Drivers (diagnostic breakdown of why users churn), and Retention Impact (transaction value at risk vs. proactive retention cost).