I design reproducible ML pipelines, structured experimentation, and practical evaluation strategies that turn data into decisions.
Focus: pipeline architecture, MLflow experiment discipline, thresholding for real-world tradeoffs.
End-to-end heart-attack risk modeling system comparing linear models and boosted trees using structured feature engineering, MLflow tracking, and precision–recall driven threshold tuning.
Tip: replace the GitHub link above with your CardioSentinel repo URL.
I help teams move from “we trained a model” to “we have a reliable ML system.”
Set up clean MLflow tracking, metrics, artifacts, and model comparison that actually guides decisions.
Turn notebooks into config-driven pipelines with disciplined preprocessing and feature governance.
Precision/recall, thresholds, calibration, and cost-aware tradeoffs tailored to your business context.
Curated write-ups based on real build-and-debug work (short, practical, and reusable).
How changing the decision threshold can improve recall when missing positives is expensive.
When to keep exploration flexible, and when to lock a workflow into a repeatable pipeline.
How to avoid leakage and keep feature engineering honest during train/test splits.