Privacy-first in-browser Generative AI web apps: offline-ready, future-proof, standards-based

What if you could run powerful AI in a web app with total user privacy, completely offline? New browser APIs make it possible.

#1about 3 minutes

A demo of client-side AI using the NPU

A computer vision application performs image classification directly in the browser without any backend calls by leveraging the device's Neural Processing Unit (NPU).

#2about 3 minutes

The case for privacy-first, on-device AI

On-device AI meets user demands for performance, privacy, and offline access while satisfying developer needs for a unified codebase and helpful abstractions.

#3about 3 minutes

Introducing the Web Neural Network (WebNN) standard

The emerging WebNN standard provides a model-agnostic, unified abstraction for near-native AI execution in the browser, designed around practical use cases.

#4about 4 minutes

Leveraging hardware like the CPU, GPU, and NPU

WebNN can access all available hardware, with the NPU offering a power-efficient alternative to the GPU for sustained AI workloads on mobile devices.

#5about 6 minutes

Getting started with the low-level WebNN API

To experiment with the emerging WebNN standard, developers must use canary browser versions and enable specific flags, but its low-level API can be complex.

#6about 7 minutes

Simplifying development with high-level AI frameworks

Frameworks like ONNX Runtime Web and Transformers.js provide higher-level, task-based abstractions over WebNN, making it easier for app developers to build AI features.

#7about 3 minutes

Best practices and the future of browser AI

Focus on user experience by providing fallbacks and progress indicators, and look ahead to upcoming built-in browser APIs like the Prompt API that abstract away model management.

#8about 2 minutes

Demo code and using web workers for performance

The demo applications are built as offline-ready Progressive Web Apps and use Web Workers to run intensive AI computations without freezing the main UI thread.