Record
E-commerce Purchase Intention MLOps
A purchase-intent model built as a reproducible workflow, with the cost of a wrong prediction kept in view.
What it is
A purchase-intent prediction record built as a reproducible local MLOps workflow. Visitor-session behavior becomes features, baseline models are compared on a holdout, and a local FastAPI surface serves the prediction so the interface can be tested rather than described.
What it proves
It shows the translation from a business question into a data problem and back again. The target is framed explicitly, preprocessing is reproducible, Logistic Regression and Random Forest baselines are compared, the tradeoffs are written down, and the prediction is connected to threshold costs instead of being presented as a score. Tests, CI, tracked reports and model-card style notes are all in the repository.
What it does not prove
It is not client work, a deployed commercial model, a live personalization system, or a tuned decision policy. There is no customer data behind it, no automated intervention, no production monitoring, and no evidence of revenue impact.
Artifacts
The public repository carries the preprocessing and training path, the local API, the tests and CI, and the model and monitoring reports with their production limits stated in the reports themselves.
Takeaway
What to take from it.
Decide what a wrong prediction costs before deciding the model is good enough.