Skip to workspace

LOCAL MACHINE LEARNING / NO API KEYS

Sort with
Evidence.

Teach a small inbox classifier. Inspect its evidence. Keep uncertain messages with a person.

Start with the six-message demo

Run the classifier, inspect “Review” items, choose each final label and approve. Your messages stay in this page’s memory. Reload clears them.

01 / Your inputs

Model scores are not calibrated probabilities. Low vocabulary coverage always requires review.

Training examples & held-out evaluation

The demo uses 24 original synthetic examples. Eight separate synthetic tests are never used to train. No real campus messages or user validation.

Run evaluation to inspect per-case results.

02 / Review every decision

Your inbox is waiting. Start with “Classify demo inbox”.

Learning uses only approved corrections, records their provenance, then invalidates the previous review. It never sends or routes a message.

03 / What this model can prove

Multinomial Naive Bayes learns word counts from labelled examples, with Laplace smoothing. Evidence compares the suggested label with the runner-up. It cannot understand intent, reliably handle languages beyond this English demo, or guarantee correct routing. Unknown vocabulary and close scores are surfaced, not hidden.

OpenAI Codex generated this new app and its synthetic fixtures under Aditya Mittal’s direction. No runtime generative model, cloud inference or user study. The proposal is a review assistant, not an autonomous helpdesk.

Algorithm background and probability caveat