Train microclassifiers in the cloud for spam detection, sentiment analysis and more.

Classr is a web app that allows users to create microclassifiers. A microclassifier is just a machine learning model (classifier) that is trained using a minimal (but still reasonable) amount of training data (4MB max).

Classr uses the bayes package at its core with custom additions for calculation of precision/recall/F1 score etc. and generation and rendering of confusion matrices.

I wanted to see what I couuld do working on something from scratch with the T3 stack while sharpening my frontend skills in the process.

Turns out, it's a super cool stack that you should definitely try out if you haven't already!

• create-t3-app
• TypeScript
• Next.js
• NextAuth.js
• Prisma
• Tailwind CSS
• tRPC

Classr trains multinomial naive bayes classifiers behind the scenes on unigrams generated by splitting strings along space characters. Train/test split is 20/80 and Laplace smoothing is at 1.

Any ML engineer with tell you that this is a very simple classifier with a lot of caveats. Use appropriately, and wield responsibly.

Classr does not support any of the following features (yet!):

• Training on n-grams.
• Tokenization based on anything but splitting documents along spaces.
• Stopword removal.
• More advanced classifiers (SVMs, random forests etc.)

This app is designed for deployment on Vercel but you might also be able to get it to work with Docker (I haven't tried this myself yet).