Private AI in the Classroom: WebAssembly Inference on the Edge

The Privacy Dilemma in EdTech

Integrating Large Language Models (LLMs) into K-12 education has historically faced one massive hurdle: student data privacy. Sending children's conversations or schoolwork to external API providers raises significant FERPA and COPPA compliance issues. Even anonymized telemetry can be problematic when the subjects are minors.

The solution? Don't send the data anywhere. Run the model directly inside the student's web browser using WebAssembly.

How WebAssembly Changes the Game

WASM allows pre-compiled C/C++ code to execute inside the browser sandbox at near-native speeds. By compiling llama.cpp to WebAssembly, we load a quantized LLM directly into the client's RAM and execute inference using the device's own CPU or WebGPU implementation.

This means network traffic drops to zero once the model is downloaded. The student can interact with the AI completely offline, and the school district never has to worry about data exfiltration.

Model Delivery and Caching

Downloading a 1.5–4 GB model file on every page load is unfeasible. We solve this with a two-layer caching strategy:

1. First visit: The model is fetched over HTTP/3 with chunked transfer encoding. A Service Worker intercepts the response and streams each chunk into IndexedDB via a cursor-based write.
2. Subsequent visits: The Service Worker serves the model directly from IndexedDB, achieving sub-100ms cold-start for the inference engine.
3. Integrity check: A SHA-256 manifest is compared on each load. If the hash diverges (new model version deployed), the cache is invalidated and a background re-fetch is triggered.

Cost Savings and Scalability

Beyond privacy, there is a tremendous economic incentive. Providing a cloud-based AI tutor to 100,000 students could cost a district tens of thousands of dollars per month in API fees. By offloading compute to the student's device, the per-query cost for the school district drops to zero.