I'm a data scientist in practice — previously a full-stack developer (Laravel & Vue.js), now focused on building end-to-end ML pipelines, working with real-world messy data, and extracting insights that actually matter.

My engineering background means I don't just build notebooks — I build structured, reproducible, deployable data pipelines.

romel = {
    "focus"      : ["Machine Learning", "Data Analysis", "Geospatial Data", "Time Series"],
    "background" : ["Full Stack Web Dev", "ERP/CRM Systems", "REST APIs"],
    "currently"  : "Building ML projects on real-world datasets",
    "strength"   : "Developer mindset applied to data science problems",
    "contact"    : "romelhasan741@gmail.com"
}

District-level crop yield prediction using satellite data + stacking ensemble

• Integrated BBS agricultural data with Google Earth Engine satellite covariates (CHIRPS rainfall, MODIS LST, EVI)
• Built a Stacking Ensemble (CatBoost + Random Forest) achieving R² = 0.8481
• Used SHAP analysis to identify district_capacity as the dominant predictive feature
• Generated an interactive Leaflet.js choropleth map for 2024 district-level predictions

Python CatBoost scikit-learn SHAP Google Earth Engine Leaflet.js

🔗 View Repository

End-to-end ML pipeline on 1.58 million crime records

• Cleaned and engineered features from 1.58M LAPD records (26 columns → structured ML-ready dataset)
• Grouped 200+ crime types into 8 categories as the classification target
• Handled severe class imbalance using SMOTE, trained with XGBoost → 69% accuracy
• Built with proper software structure: src/, notebooks/, tests/, main.py

Python XGBoost SMOTE pandas scikit-learn imbalanced-learn

🔗 View Repository

Data Science & ML

Web Development (Background)

Tools & Infrastructure