Session

404: Backend Not Found, but AI survived

I wanted to find out how much of an AI app can run in the browser with no backend, so I built a live demo for every layer I could: Chrome's built-in Gemini Nano through the Prompt API, TensorFlow.js and MediaPipe for vision, WebGPU and WebAssembly for compute, and Transformers.js, WebLLM and ONNX Runtime Web for bigger models, with OPFS holding the weights. A lot of it worked. A lot of it broke in ways no tutorial warned me about.

This talk is the breakage log, one failure at a time. WebLLM's model weights redirect from Hugging Face to a CDN and fail the browser's cross-origin check. The moment the page sets the COOP and COEP headers that SharedArrayBuffer needs, every asset from another origin gets blocked (Pyodide from jsDelivr, MediaPipe models from Google Cloud Storage, MobileNet from tfhub.dev), and without those headers, multithreaded WebAssembly quietly falls back to one thread. A TensorFlow.js model that fetched fine and then died with "Cannot read properties of undefined (reading 'producer')" because it was the wrong model format. Gemini Nano, which is unavailable without the right Chrome flags and a one-time model download, and whose API name has changed along the way. And the UI problem nobody mentions: a model that downloads silently looks exactly like a broken app.

You will see each failure live, with the cause and the fix, and with the WiFi off at the end, a demo or two might not survive the room. You will leave with a field guide to what is actually safe to ship in the browser today, a decision framework for on-device, server or hybrid AI, and the fixes for the traps you would otherwise hit on day one.

Aprajita Verma

Frontend Architect

Delhi, India

Actions

Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.

Jump to top