Dataset Drift Diff

Build a small self-serve tool for ML engineers and indie teams that compares two CSVs or JSONL datasets before and after a pipeline change, then produces a plain-English “what silently changed?” report: missing columns, shifted label balance, duplicate-rate changes, outlier changes, PII-like fields introduced, and example records worth reviewing. The unusual angle is not generic data profiling—it is a release-diff tool for datasets, like a git diff for training data. Start with file uploads, local/browser-side processing where possible, and downloadable HTML reports. Charge for scheduled cloud-source checks (S3/GCS/GitHub), Slack alerts, saved baselines, and team history. It fits one-hour-per-day maintenance because the core workflow is asynchronous and self-serve. Day 1 MVP: upload two CSV files, calculate schema/null/distribution differences with pandas or Polars, and generate a single report page. Free implementation stack: GitHub Pages or Cloudflare Pages, Python/FastAPI on Render or Fly.io free tier, SQLite, and open-source profiling libraries. Avoid claiming full compliance; position PII detection as a review hint.