Power BIPlotlyJupyterColabLangChainOllamaClaude APIPostgreSQLSQLiten8nAzureAWS
95.6%failure recall, predictive maintenance
−33%operating cost from a tuned threshold
~100Kreal orders modeled, Olist capstone
99%evaluation score, NTI data analytics
About
// who is behind the code
I'm a Data Science and AI student with two years of hands-on experience turning raw datasets into decision-ready insights with Python, SQL and Power BI.
I finished an eight-month Microsoft and DEPI Machine Learning Engineer program covering data pipelines and Azure, and a data analyst internship at the National Telecommunication Institute with a 99% evaluation. AWS and NVIDIA certified.
Alexandria, EgyptB.Sc. 2023 to 2027Arabic and English
about.py
# about.pyengineer={"name":"Marwan Mohamed Saeed","role":"Data Science & AI","focus":["ML pipelines","RAG apps","Power BI"],"stack":["Python","SQL","PyTorch","Docker"],"open_to":"junior Data / AI roles",}defnext_step(you):return"let's build it together">>> next_step("you")'let's build it together'
Selected work
// six projects, code on GitHub
Team project · 6 members
Predictive Maintenance: machine health monitoring
An end-to-end system that scores failure risk on a milling process, names the likely failure mode, explains each alert with SHAP and recommends a maintenance action. The decision threshold comes from a business cost matrix, not the default 0.5.
0.886PR-AUC, tuned Random Forest
0.978ROC-AUC
3 / 68failures missed on the sealed test set
127automated tests
Cost-optimized threshold3 missed of 68
Default 0.5 threshold12 missed of 68
Each square is one real machine failure in the 2,000-reading test set. Red squares are failures the model did not flag. Same model, different threshold.
Streamlit app with six pages, including a grounded RAG assistant, drift monitoring and a maintenance scheduling simulator. Packaged with Docker. Trained on the synthetic AI4I 2020 benchmark, so it is a prototype for learning and not a safety system.
Chat with lecture PDFs and get answers with page citations, then generate practice quizzes. A seven-step validation pipeline, including a blind solve, removes questions with wrong answer keys. Runs locally with Ollama or with the Claude API.
One shared data pipeline feeds five models: return and late-delivery classification, review-score and revenue regression, and customer segmentation with K-Means, DBSCAN and agglomerative clustering. Tuned models are exported and served in a Streamlit app.
Retail intelligence dashboard on a fact and dimension model with DAX measures. It covers KPIs by membership tier, customer insights by income, education and region, and product profit and return analysis.
Streamlit app comparing Logistic Regression and Random Forest, with recall and F1 prioritized because a missed cancer costs more than a false alarm. Includes permutation importance, ROC and precision-recall curves, and 5-fold stratified cross-validation.
I'm looking for a junior Data or AI role, and I take freelance data projects. Send me the problem and the data you have, and I'll reply with how I would approach it.
