I got tired of waiting 500ms every time I typed something into my app's command bar.

When you just want to type "Algorithms exam tomorrow at 9am" and hit Enter, sending a request to a cloud LLM feels broken. You feel the network lag, it fails on bad campus Wi-Fi, and paying for API calls just to figure out if a sentence is a calendar event or a to-do item makes no sense.

So for Stedy
, I ripped the cloud call out of the hot path and moved classification into the browser.

The setup is pretty simple:

Deterministic TypeScript code extracts dates, time ranges, and known course names.
ADS-1.5M, a small INT8-quantized model (~1.5M parameters), runs in WebAssembly via onnxruntime-web/wasm to classify the intent (EVENT, TASK, or NOTE) and category.

Once the .int8.onnx file and vocab.json are cached in CacheStorage, a forward pass takes less than 10ms locally:

import * as ort from "onnxruntime-web/wasm"; const session = await ort.InferenceSession.create( "/models/intent-classifier.int8.onnx", { executionProviders: ["wasm"] } ); const feeds = { input_ids: new ort.Tensor("int64", tokenIds, [1, tokenIds.length]), }; const results = await session.run(feeds); 

Is a 1.5M model as smart as a frontier LLM? Obviously not. If the local confidence score is too low, it still falls back to an API call.

But for 90% of daily inputs, running a tiny quantized model in WASM makes the whole app feel instant, works completely offline, and keeps user notes on their own device.

If you're curious to test it out, it's live inside Stedy