Principal Data Scientist — Forecasting, Optimization & Recommender Systems
Retail & consumer · Clean energy & industrial operations
Resume · Portfolio · Certificates
I build machine learning systems that change operational decisions — what to stock, where to put a warehouse, how to run a plant on less energy. Twelve-plus enterprise engagements, most of them ending in something running in production rather than a slide deck.
- 🏭 23% lower natural-gas consumption at a manufacturing plant — real-time recommender on live sensor data, with automated quality safeguards that kept batch failures at zero
- 🚚 18% lower total supply-network cost — warehouse and hub placement, milk-run routing, fleet sizing, benchmarked against the incumbent network
- 📦 Demand forecasting error cut from 300% → 20% at site level; inventory replenishment that reduced carrying cost and stock-outs at the same time
- 🎯 Customer targeting precision lifted 3% → 25%; churn predicted at 79% precision across 100GB+ of point-of-sale data
- 🧱 Founded and scaled a data science team 1 → 10; applied ML lead for an internal MLOps platform now running ~25 production models
🔭 Currently Principal Data Scientist at Synapse Analytics
🎓 PhD in Computer Science & MSc in Bioinformatics, Nanyang Technological University, Singapore
🌍 Based in Cairo — open to relocation (Europe, China, Japan)
Demand & operations. Time-series forecasting, inventory and replenishment planning, constrained production scheduling. The interesting part is rarely the model — it's that over- and under-forecasting cost different amounts, and the loss function has to know that.
Optimization on physical networks. Facility location, vehicle routing, fleet sizing, capacity planning. Clustering to decompose the problem, routing to solve within it, real road distances rather than straight lines, and a cost curve rather than a single "optimal" answer so the people who own the network can make the trade-off themselves.
Energy & sustainability. Reducing consumption in industrial processes without compromising product quality. My best result here came from choosing an interpretable method over a more accurate black box — process engineers won't act on a number they can't interrogate, so a recommendation nobody trusts saves nothing.
Personalization & ranking. Recommender systems, lead scoring, propensity and churn modelling — usually delivered as a ranked worklist written back into the system people already use, not a dashboard they have to interpret.
Applied LLMs. Extraction pipelines, RAG, agents, and small-model distillation — mostly in service of making unstructured operational data usable, and making per-item inference cheap enough to actually run at volume.
ropensci/dbparser — R package · 50K+ downloads
Open-source package that parses and integrates heterogeneous public databases into one unified object model, so downstream analysis is written once regardless of source. Peer-reviewed and hosted by rOpenSci. I co-developed it and set the research direction, feature priorities and testing strategy. The lesson generalises: the highest-leverage work is often the integration layer nobody wants to own.
Chemogenomic-DTI-Prediction-Methods — ⭐ 48 · 🍴 14
Companion resource to my survey paper on predicting interactions between two entity sets from a sparse, severely imbalanced interaction matrix — structurally the same problem as implicit-feedback recommendation and cold-start ranking. Organised links to the field's methods, source code and feature-extraction tooling.
Life-with-Recs — recommender systems
Working notes and implementations on ranking and recommendation — collaborative filtering, matrix factorization, evaluation methodology.
Life-with-Graphs — networks
Analysis and prediction on graphs. Relevant to anything shaped like a network: supply chains, routing, entity resolution, knowledge graphs for retrieval.
Life-with-LLMs — GenAI
LLMs, RAG and agents. Where I test things before they earn their way into production work.
Life-with-NLP — NLP
Text modelling — classification, extraction, embeddings, semantic search.
Also here — competitions, scrapers and tooling
CAFA-6-Protein-Function-Prediction— Kaggle extreme multi-label classification (26K labels). EDA and feature-engineering pipeline.bggscraper— scraping toolkit built on the BoardGameGeek XML API. Data acquisition when no dataset exists is a skill in itself; I've used the same approach to build a used-car pricing dataset from classifieds when the market had none.heroic_surfer— scraped corpus feeding an LLM-powered Shiny app for exploring a self-development content library.my_r_snippets— R snippets. Free to copy.
4 peer-reviewed papers and 1 book chapter · Best Paper Award, InCoB 2016
- Computational Prediction of Drug–Target Interactions using Chemogenomic Approaches: An Empirical Survey — Briefings in Bioinformatics, 2018
- Drug–Target Interaction Prediction with Graph Regularized Matrix Factorization — IEEE/ACM TCBB, 2017
- Drug–Target Interaction Prediction using Ensemble Learning and Dimensionality Reduction — Methods (Elsevier), 2017
- Drug–Target Interaction Prediction via Class Imbalance-Aware Ensemble Learning — BMC Bioinformatics, 2016
Full list on ResearchGate.
Python R SQL MATLAB · PyTorch scikit-learn XGBoost Transformers pandas NumPy
Docker Git MLflow Spark AWS Azure Neo4j · Shiny Plotly ggplot2 Leaflet
I'm most interested in retail and consumer businesses with genuinely hard demand, pricing and inventory problems — and in the clean energy and sustainability side of industrial operations, where making a better decision and cutting emissions turn out to be the same problem.
- LinkedIn: linkedin.com/in/aliezzat1985
- Email: aliezzat1985 AT gmail DOT com
- ResearchGate: Ali Ezzat
🎲 Board games · 📚 reading · 🥋 martial arts