A collection of production-grade feature engineering skills for applied data science use cases. Each skill is a self-contained, prescriptive workflow covering problem framing, schema mapping, leakage hygiene, feature construction, encoding, validation, and model handoff.
End-to-end feature engineering workflow for predicting which customers will stop engaging with a platform. Built for food-delivery but applicable to any consumer app with transaction and session data. Covers:
- Panel and label construction with right-censoring
- 12 feature families — tenure, recency, frequency/cadence, monetary, trends, habit/breadth/location, delivery health, support signals, promotions, cohort, marketing engagement, and restaurant context
- 18-trap leakage catalogue specific to food-delivery data (mutable snapshots, late-settling refunds, ML backfills, and more)
- Categorical encoding decision rules with time-aware target encoding for multi-cutoff training sets
- Model-agnostic feature filtering via correlation deduplication and mutual information ranking
- Class imbalance handling and evaluation metric guidance
- 24 reusable pandas helpers in
scripts/inactivity_features.py
This project was built using an AI-augmented workflow with Claude. The division of work was deliberate:
My role — systems architect and lead QA: I defined the problem scope, made every substantive domain decision (which features matter, which leakage traps are real - for example in the food delivery industry, what a production inactivity model actually needs), reviewed every generated output critically, and drove the iterative refinement through multiple review cycles.
Claude's role — implementation accelerator: Claude handled rapid boilerplate generation, syntax, code structure, and cross-file consistency checks — the mechanical work that would otherwise slow down translating domain knowledge into a usable artifact.
The result is a skill that reflects professional experience in applied data science rather than a generic template. The domain-specific depth — food-delivery leakage traps, settlement buffer patterns, preference-anchored recency, time-aware encoding for multi-cutoff panels — comes from applying that experience deliberately, with AI as the implementation layer rather than the decision-maker.