A model, deployed.

The original project was a data science final: predict depression among students from survey responses using 18 behavioral and academic features. The model trained well — 82% accuracy, AUC of 0.91 — but it existed only as a Scikit-learn pipeline inside a notebook.

That’s a common stopping point. But a model that only runs locally isn’t a product. The interesting question was: what does it take to make it callable by anyone?

The stack is deliberately simple. FastAPI for the HTTP layer — it handles request validation, OpenAPI generation, and async routing without ceremony. Pydantic schemas enforce input contracts at the boundary: wrong types get rejected with a 422 before the model ever sees the data.

Everything runs in a single Docker container. The image includes the serialized model artifacts, the preprocessing pipeline, and a /health endpoint so the orchestrator knows when the service is ready. Render pulls the image on every push to main, rebuilds, and swaps traffic with zero downtime.

CI/CD via GitHub Actions runs a smoke test on every pull request. The deploy only happens if the container starts and the health check passes.

SHAP analysis on the full 18-feature model surfaced a clear hierarchy. The five features below explain the majority of prediction variance — the rest added noise rather than signal. Trimming to them cut inference payload size without measurably affecting accuracy.

The model also takes Age as a sixth input — a weaker individual predictor but one that sharpens confidence at boundary cases.

The Render deployment runs on the free tier — which means the instance sleeps after a period of inactivity. The first request after a cold start incurs a spin-up delay of up to 90 seconds. Subsequent requests hit the <500 ms target consistently.

This is a deliberate tradeoff: zero infrastructure cost for an unpredictable audience, at the expense of cold-start latency. A production deployment would use a paid instance or a keep-alive ping to hold the process warm.

The training data is synthetic — generated to approximate the distribution of a real student mental health survey but not drawn from clinical subjects. The model should not be used to make real-world clinical assessments.

The API is also unauthenticated. For a real deployment handling sensitive health data, you would want JWT or API key gating at minimum, and likely compliance review depending on jurisdiction.

SHAP feature selection was done on the training split. Selection bias is possible — the dropped features may generalize better on out-of-distribution data than the retained ones.