Senior Data Scientist · customer value management, churn and experimentation · Power BI and Microsoft Fabric
I build the systems that decide which customer gets which offer, and prove each one against a holdout. Five years across fitness, healthcare diagnostics, telecom and banking, shipped as governed data products. Currently Senior Data Scientist at Leejam Sports (Saudi Arabia), working remotely from Cairo.
- Retention engine: 260 personalised-offer programs, each judged on incremental lift against a control group; about SAR 20M a month.
- Member 360 on Microsoft Fabric: 80+ features behind churn scoring, segmentation, next-best-offer and target/control assignment.
- ML segmentation at IDH: replaced rule-based campaign lists; monthly campaign impact from about EGP 10M to EGP 40M in three months.
Employer work stays private. The projects below are public, built on real open data, and generated from code you can read.
Six end-to-end projects on real public data. Each runs the full cycle in notebooks (data audit and EDA, out-of-time validation against baselines and a tuned challenger, calibration, SHAP, the business decision and a model card), then ships a Power BI report with its own design. Every published number is reproduced and asserted in code; unit tests run in CI.
| Project | What the notebooks do | What they found | |
|---|---|---|---|
| CVM Command Centre | Churn model, A/B readout, uplift and policy evaluation | Champion/challenger: a regularised logistic model beats the production churn model (AUC 0.770 vs 0.761), and personalising the e-mail adds nothing measurable. | notebooks |
| Telecom Churn & Subscriber Value | Churn model, calibration for over-sampled data, retention economics | On a realistic 2% monthly churn base the default $25 offer loses about $80K; a $5 offer aimed at positive expected value pays. | notebooks |
| Card Portfolio & Risk | WoE scorecard vs monotone GBM, fairness, fraud model tested forward in time | A monotone LightGBM ties the production PD model (0.778 vs 0.776) and passes the four-fifths fairness test; an unweighted fraud model lifts PR AUC from 0.74 to 0.79. | notebooks |
| Retail Network Performance | Rolling-origin forecast with intervals, fixed-effects promotions, competitor event study | With fixed effects, promotions add 39% (95% CI 37-41%); a competitor costs 16% of sales within 500 m and under 2% beyond 5 km. | notebooks |
| Retail Bank Campaign Strategy | Propensity model, random vs out-of-time validation, contact-policy economics | Random cross-validation claims AUC 0.80; on later months it is 0.72. Keep the call cap and use the score to order calls, not to cut clients. | notebooks |
| Olist Marketplace Performance | Late-delivery risk at checkout, drift diagnosis, promise-padding policy | The first late-delivery model collapsed when promise rules changed (AUC 0.60); a rule-plus-model blend holds 0.76, and a 3-day pad on the riskiest 10% avoids 12% of late orders. | notebooks |
The semantic models (TMDL) and reports (PBIR) are generated by Python, so every measure and visual is reviewable in a diff.
- Start from the decision. Every model answers who to contact, what to offer, or where to open.
- Prove it against a holdout. Campaigns run as A/B tests; results are judged on incremental lift.
- Define it once. KPIs live in one governed semantic layer that reports, models and AI assistants all share.
Modelling: Python, scikit-learn, LightGBM, Optuna, SHAP, statsmodels; uplift and causal inference (DiD, fixed effects), A/B testing, scorecards, forecasting, R, SQL
BI and data: Power BI (DAX, calculation groups, field parameters, PBIP/TMDL), Microsoft Fabric, Power Query, SQL Server
AI: Amazon Bedrock knowledge bases and Claude-based assistants grounded in governed KPIs