foxypandas
Data Scientist·ML Engineer
Building ML systems from data exploration to deployment — models with a measurable business outcome, shipped as reproducible code and services.
Selected projects
04Steel Temperature Prediction
Predicting final alloy temperature in ladle processing so a steel plant can cut heating energy. Event-level data from seven sources aggregated per batch; Ridge beat tuned LightGBM and RF on CV; stability confirmed with bootstrap and Wilcoxon, features explained with SHAP.
Disaster Tweets
Systematic comparison of transformer fine-tuning strategies for binary tweet classification: DistilBERT → BERT-base, heads, schedules, ensembling. Honest score without the leaked original dataset that inflates the top of the leaderboard.
Uplift Modeling — X5 RetailHero
Who should actually get the SMS? Uplift models on retail A/B data estimate the incremental effect of a message per customer; SHAP explains who responds; a profit model turns scores into a targeting policy. Served as a FastAPI endpoint in Docker with MLflow tracking.
Oil Well Region Selection
Choosing where to drill: reserve prediction per well, selection of the best sites, and a bootstrap profit distribution per region. Picks the only region that satisfies the business risk threshold instead of the one with the highest point estimate.
Stack
used in the projects aboveModeling
- scikit-learn
- CatBoost · LightGBM
- PyTorch · HuggingFace
- Uplift · causal inference
- Time series
Data & analysis
- Python · pandas · Polars
- SQL
- Bootstrap · A/B tests
- SHAP
- Matplotlib · Plotly
Engineering
- FastAPI · Pydantic
- Docker · Compose
- MLflow
- Git · pytest
- Airflow
Practice
- Leak-free validation
- Confidence intervals
- Model cards
- Reproducible pipelines
- Monitoring
Contact
open to DS / MLE rolesI like problems where the model is only half of the answer — the other half is the decision it drives and the service that delivers it.
Happy to talk about uplift and causal ML, NLP, industrial data, or the practical side of shipping models.