Small models in the browser

Each app runs a 0.8B language model in your browser on WebGPU. There is no server. All data is synthetic. You need desktop Chrome or Edge with WebGPU and about 2 GB of free GPU memory.

Four complex apps with a quiet assistant (no chat)

1. Open an app. Wait until the panel on the right stops showing "Loading the model" (about 2 minutes on the first visit).
2. Click an item in the list, for example the first claim.
3. Read the assistant's cards, or click "What's next", "Summarize" or "Insight". Decisions stay yours; Undo reverts the assistant.

• Claims: coverage, fraud checks, settlement within an authority limit.
• Purchase-to-Pay: invoices, three-way match, approvals, payment run.
• Clinic front office: referrals, eligibility, booking, billing. The model never saw this app in training.
• Underwriting Workbench: 25 loan files, 11 tabs, a 9-step workflow.
• Onboarding: new hires and five setup forms each. The assistant reads the offer and fills the forms (built on the surfaces, not in the training data).