This project: batch
File or schedule driven: ingest → validate → score → hand off CSV (or parquet) to CRM, analytics, or compliance. Same pattern as many production churn or risk batch jobs.
Skip to content
Portfolio · ML systems
Batch Scoring
Batch scoring for real pipelines.
Turn a customer CSV into a scored file: validation, training-aligned features, predict_proba, versioned output. Built for jobs that look like ETL — not microsecond APIs.
The demo below uses the API when you run uvicorn web.main:app from the repo. A static mirror of this page has no backend — use the CLI or CI there.
Tests · pytest suite
CLI · flags & exit codes
Optional · run .meta.json
Try the pipeline
Upload a Telco-style customer batch (same shape as the input contract below), or run on a small built-in sample — no CSV on your machine required. The server validates, preprocesses, and scores; then you can download the result.
Scoring uses the small E2E fixture model committed in this repo (under tests/fixtures/) so the full path is real and testable — it is not a production or client-specific model. Run with sample batch loads five Telco-style rows from the same fixture set.
Positioning
Contrast: online API
Request/response, SLOs, autoscaling. Valuable, but a different skill slice. This repo stays intentionally narrow: offline orchestration and clean batch output.
End-to-end flow
InRaw CSV
GateValidate
PrepPreprocess
ScoreModel
OutScored CSV
Stack
Python
pandas
scikit-learn
joblib
PyYAML
pytest
I/O contract
Input (Telco-style)
customerID, tenure, Contract, MonthlyCharges, TotalCharges, PhoneService, InternetService, …
Output
customer_id churn_score predicted_label model_version scoring_timestamp
Output review (Jupyter)
After you run the pipeline locally, the repo includes a notebook that validates the scored CSV (schema, row counts, probabilities in
View notebook on GitHub → · readable in the browser; clone the repo to execute cells.
[0,1]) and plots the churn-score distribution.
View notebook on GitHub → · readable in the browser; clone the repo to execute cells.
Interview pitch (30 seconds)
- Problem: After training, a lot of value is in recurring batch scores — not every product needs a real-time endpoint.
- What I shipped: One pipeline: strict input checks, same feature engineering as training, probabilities and labels, plus model version and timestamp on every row.
- Ops detail: CLI for cron/CI, non-zero exit on failure, optional JSON manifest so you know which input and threshold produced a file.