Mohamed Ayman, Senior Data Scientist

Mohamed Ayman

Senior Data Scientist · Cairo, Egypt

I build the systems that decide who gets which offer, and prove each one against a holdout.

I'm a Senior Data Scientist at Leejam Sports, a Saudi fitness-centre chain, working remotely from Cairo. Five years of churn, segmentation, and campaign measurement in fitness, healthcare diagnostics, and telecom, shipped as governed data products on Microsoft Fabric.

Email meDownload the CV (PDF)

Treated group against a holdoutIllustration, not real data: how each campaign is judged. The gap between the treated group and the holdout is the incremental lift.Offer goes outIncremental liftTreatedHoldoutTime
Illustration, not real data: how each campaign is judged. The gap between the treated group and the holdout is the incremental lift.

Selected work

Three employer projects, each measured against a holdout or a business decision. The data and code stay private; each write-up covers the problem, the method and the reported result.

  1. 01

    ~SAR 20Ma month

    LeadConversionActiveGrowthExpiringRetentionLapsedReactivationwin back lapsed membersEvery program: offer group vs holdoutOffer groupHoldoutJudged on incremental lift, not response

    Leejam Sports, a fitness-centre chain in Saudi Arabia

    A lifecycle retention engine: 260 personalised-offer programs

    Senior Data Scientist2025 to now

    Personalised offers for every stage of the member lifecycle, each program judged on incremental lift against a holdout group.

    • Four program families: conversion, growth, retention before expiry and reactivation
    • Member 360 on Microsoft Fabric: 80+ features behind every model and program
    • Target/control assignment for every program, judged on incremental lift

    260 programs live, generating about SAR 20M (about USD 5.3M) a month.

    • Microsoft Fabric
    • Python
    • SQL
    • Segmentation
    • A/B testing
    Read the case study
  2. 02

    ~4xmonthly campaign impact

    1040Campaign impact, EGP M a month1545Visits a month, thousandsRule-based listsML segments, 3 months later

    IDH (Al Borg, Al Mokhtabar), 600+ diagnostic-lab branches in Egypt

    Replacing rule-based campaign lists with ML segmentation

    Data Scientist Supervisor2024 to 2026

    Customer models on visit behaviour, test profiles, and payment behaviour, a customised offer for each segment, and an A/B test before any offer was scaled.

    • Customer models on visit behaviour, test profiles and payment behaviour
    • A customised offer per segment, A/B-tested against a control before scaling
    • Automated impact reporting; supervised the junior data scientists

    Monthly campaign impact grew from about EGP 10M to EGP 40M, and visits from 15K to 45K, in 3 months.

    • Python
    • ML segmentation
    • Oracle SQL
    • Power BI
    • A/B testing
    Read the case study
  3. 03

    15 of 20recommended sites accepted

    Accepted (15)Not taken (5)

    IDH, healthcare diagnostics in Egypt

    Choosing where to open new lab branches with a geospatial model

    Data Scientist Supervisor2024 to 2026

    A scoring model that ranks candidate locations, replacing manual feasibility studies with a shortlist the expansion team can act on.

    • Scored candidates on branch performance, customer behaviour, competitor and facility density, population and GDP
    • Turned the scores into a ranked shortlist for the expansion team
    • The model recommends; people decide

    15 of the first 20 recommended sites were accepted.

    • Python
    • Geospatial features
    • Scoring model
    Read the case study

Also at work

  • KPI-grounded analytics assistantLeejam

    Owned the Amazon Bedrock Knowledge Base that grounds a Claude-based assistant in Microsoft Teams in schemas, KPI definitions, business rules, guardrails, and numeric validation.

  • Governed semantic modelLeejam

    A galaxy-schema Fabric model unifying 5+ source systems into leadership KPIs; semantic-model maintenance time cut 40%.

  • CX text analyticsLeejam

    Sentiment and intent classification on support tickets and survey text, tying experience drivers to specific centres.

  • Executive reporting automationHuawei

    Rebuilt executive reporting in Power Query and Power BI: about 4 hours to under 1 minute per cycle, 30+ hours saved a month.

Data science case studies

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, explainability, and the business decision with a model card. Each then ships a Power BI report with its own design. Every number is reproduced from code.

  • EDA and leakage audit
  • Out-of-time validation
  • Champion vs challenger
  • Calibration and SHAP
  • Decision and monitoring
  • Tests and CI
  • CVM Command Centre: first report page

    Retention · churn · offer experiments

    CVM Command Centre

    Base dynamics, cohort retention, an out-of-time churn model, an A/B test readout with p-values computed in DAX, customised offers by segment, and an uplift-targeting simulator.

    The uplift model's top 10% returns 3.6x the revenue of a random 10%.

    NotebooksChampion/challenger: a regularised logistic model beats the production churn model (AUC 0.770 vs 0.761), and personalising the e-mail adds nothing measurable.

    Power BI (PBIR, TMDL, DAX UDFs), Python, scikit-learn

  • Telecom Churn & Subscriber Value: first report page

    Telecom · CVM

    Telecom Churn & Subscriber Value

    Churn drivers across tenure, handset, network and care for 51K real subscribers, a churn model, and a retention-economics simulator with four what-if inputs.

    Ranking offers by revenue at risk reaches 58% of lost revenue with 30% of offers, vs 44% ranking by churn risk.

    NotebooksOn a realistic 2% monthly churn base the default $25 offer loses about $80K; a $5 offer aimed at positive expected value pays.

    Power BI, DAX what-if, Python, gradient boosting

  • Card Portfolio & Risk: first report page

    Payments · credit risk · fraud

    Card Portfolio & Risk

    The analysis a card network's advisory team builds for an issuer: portfolio health, roll rates, a behavioural PD scorecard, a credit-line policy simulator and a fraud alert-budget simulator.

    Reviewing 0.1% of transactions stops 76% of fraud at 83% precision, tested forward in time.

    NotebooksA 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.

    Power BI, DAX, Python, scorecard scaling

  • Retail Network Performance: first report page

    Retail · store operations

    Retail Network Performance

    Like-for-like growth, a store league table, promotion and competitor effects, staffing patterns and a six-week sales forecast for 1,115 stores, with a slide-in filter panel on every page.

    A new competitor nearby costs a typical store 7% of sales; the forecast cuts error 28% vs a naive baseline.

    NotebooksWith fixed effects, promotions add 39% (95% CI 37-41%); a competitor costs 16% of sales within 500 m and under 2% beyond 5 km.

    Power BI (bookmarks, filter panel), DAX, Python

  • Retail Bank Campaign Strategy: first report page

    Banking · direct marketing

    Retail Bank Campaign Strategy

    Segment lift through a field parameter, a contact-strategy simulator built on a what-if parameter, and the macro backdrop behind 41K real telemarketing contacts.

    Capping calls at five per client cuts 15% of calls and keeps 96% of subscriptions.

    NotebooksRandom 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.

    Power BI, field parameters, DAX UDFs

  • Olist Marketplace Performance: first report page

    E-commerce · marketplace

    Olist Marketplace Performance

    An executive report on 100K real marketplace orders: growth, delivery and reviews, cohort retention and seller concentration, with a calculation group for period comparison.

    A late order costs two review stars: 4.3 on time, 2.3 when late.

    NotebooksThe 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.

    Power BI, calculation groups, DAX UDFs

What clients say

5.0★★★★★

  • 39 public reviews
  • 100% Job Success
  • 58 Upwork jobs
  • Top Rated
See every review on Upwork
  • ★★★★★
    I think Mohammed is one of the top candidates in his field. Highly recommended.
    Upwork client · Data analysis
  • ★★★★★
    Completed work ahead of deadline and proactively uncovered a few issues in the data set that saved us time later.
    Upwork client · Repeat client, survey research

Verbatim public reviews from my Upwork profile; rating and counts as of August 2026.

Experience

Professional roles, newest first. Updated November 2026.

  1. 2025

    Senior Data Scientist

    Leejam Sports · Saudi Arabia · remote from CairoApr 2025 – present

    • Built a 260-program retention engine, now generating about SAR 20M a month.
    • Member 360 on Microsoft Fabric: 80+ features behind churn scoring, segmentation, next-best-offer, and target/control assignment.
    • Owned the Amazon Bedrock Knowledge Base grounding a Claude-based analytics assistant in Microsoft Teams; cut semantic-model maintenance 40%.
  2. 2024

    Data Scientist Supervisor

    IDH (Al Borg, Al Mokhtabar) · EgyptMay 2024 – May 2026 · consultant from Apr 2025

    • Replaced rule-based campaign lists with ML segmentation: monthly impact from about EGP 10M to EGP 40M in 3 months.
    • Geospatial site-selection model: 15 of the first 20 recommended branches accepted.
    • House Call propensity model: about EGP 3M a month in incremental revenue. Supervised junior data scientists.
  3. 2023

    Data Analytics Officer

    Credit Agricole Egypt · CairoJul 2023 – Apr 2024

    • Automated recurring reporting in SQL, SAS, and R, raising reporting efficiency 40%, and built Power BI dashboards for retail leadership.
  4. 2022

    Data Analyst

    Huawei Technologies · CairoJan 2022 – Jun 2023

    • Executive reporting rebuilt in Power Query and Power BI: about 4 hours to under 1 minute per cycle, 30+ hours saved a month.
    • Owned dealer and operator incentive calculations, and set regional sales targets aligned to stock.
  5. 2019

    Freelance Statistician & Power BI Consultant

    Upwork and private clients · Remote · Saudi Arabia, Oman, UAE, Egypt, Malaysia, USOct 2019 – present

    • 120+ projects; Top Rated with 100% Job Success across 58 Upwork jobs. One researcher has returned for 11 projects.
    • Thesis and research statistics for master's and PhD researchers in Arabic and English: SEM, EFA and CFA (AMOS, lavaan), mixed-effects models in R, panel econometrics in Stata, and NVivo qualitative analysis.
    • Eight survey waves (2024–2026) for an international research agency: a Python-to-SPSS pipeline that labels raw panel exports, builds banner crosstabs, and feeds a Power BI dashboard.
    • Power BI for an international health organisation's country office: a 9-page emergency-operations dashboard covering response and preparedness, with its own design system.
    • Monthly budget-vs-actual reporting for a multi-brand investment group, from a data-quality audit to an Arabic executive deck and a BI roadmap.
    • Led delivery of a full-stack AI interior-design MVP for a Gulf technology company: Spring Boot API, React front end, and a Python data and image-generation pipeline.

Education

  • M.Sc. in Data ScienceFaculty of Graduate Studies for Statistical Research, Cairo University · In progress · expected 2027
  • B.Sc. in StatisticsFaculty of Economics and Political Science, Cairo University · 2021

Next: Data & AI Engineer on Microsoft Fabric Production ML, MLOps, and LLM systems grounded in governed data. DP-600 Fabric Analytics Engineer exam planned for 2027.

How I work

Five habits behind every project above.

  • Start from the decision

    Every model answers a business question: who to contact, what to offer, where to open. If no decision changes, the model isn't needed.

  • Prove it against a holdout

    Campaigns run as A/B tests against control groups, so results are judged on incremental lift, not on who happened to respond.

  • Define it once

    Features and KPIs live in one governed layer, so reports, models, and AI assistants all use the same numbers.

  • Keep people in the decision

    Models rank and recommend; the people who own the outcome decide.

  • Ground AI in governed data

    An assistant's KPI answers are checked against the semantic model, not taken on trust.