The classification lab Experimental
Your words.
A real decision.
Try a small text encoder on a banking support message. MemCat compares its meaning with 77 categories and flags uncertain matches for review.
Load once. Classify locally.
First load downloads about 23 MB of model weights, a 10 MB category index, and runtime assets. Your browser may cache them for later.
Write a message.
Or try an example
This demo suggests a support category. It takes no banking or financial action.
Inspect the decision.
A little context. A clearer route.
Load the model, then classify your own message. The category, alternative and measured time will appear here.
Evaluate your own labelled examples Developer tools
A prediction is only useful if it is right.
Supply 1–100 messages with known categories. The SDK reports correct and incorrect automatic routes, coverage, review rate and per-category precision/recall. These are results on your examples, not a general accuracy claim.
Use a unique id, text and expectedLabel for each example. Use null when the correct handling is review. Maximum 400 KB of UTF-8 JSON. The downloaded report includes the supplied text.
A run will appear here. No estimated savings, invented costs or provider confidence scores.
What this run tells you.
A real encoder, locally. MiniLM-L6-v2, quantized to q8, runs with one WebAssembly thread. MemCat compares normalized embeddings with prepared category examples.
A narrow task. These 77 categories cover banking support. Mixed requests, unrelated topics and ambiguous language can still be misclassified.
A measured run. The timer includes token validation, embedding and routing. Setup is separate. Zero API calls does not mean zero device compute, download or energy cost.