Early-diabetes symptoms are easy to dismiss and awkward to check: booking a lab visit for a vague suspicion feels like overkill, while symptom-checker sites either demand sign-ups or ship your health answers to a server. The bar was a check that is instant, free, and private by construction — not by promise.

No backend budget and no appetite for holding health data — so inference had to run 100% client-side. The model artifact had to be small enough to ship with a static page, and the whole pipeline had to be reproducible: anyone should be able to retrain from scratch and get the identical artifact.

Trained and compared 10 algorithms plus a 5-expert stacking ensemble (train/compare.py), publishing the full comparison table in the repo instead of cherry-picking the winner. A CTGAN augmentation study tested whether synthetic tabular data helps on this small clinical-style dataset — with honest, mixed results kept in the write-up rather than buried.

The trained model is exported to a JSON artifact and ported to TypeScript, with a parity fixture + test gate proving the browser scores exactly what Python scored. A verify_all.sh pre-deploy gate (retrain → export → parity → bundle → hashed build) means the shipped site can never drift from the evaluated model.

Live static site: 16 dropdown questions → instant local risk pattern, with methodology, benchmark, and GAN-study pages alongside it. Zero inference cost, zero data liability, fully reproducible from one commands block in the README.

Add probability calibration (Platt/isotonic) and surface calibrated bands instead of a raw score; run a proper fairness slice across age/sex subgroups before calling it a screening aid; and package the flow as an installable PWA for low-connectivity clinics.