How Interfaze Runs an AI Service on Vercel

Interfaze builds models for tasks such as reading documents and finding objects in images. The company wants users to check results, not simply accept an answer. For example, its model can return the total from an invoice. It also shows the area of the page containing that number. A confidence score indicates how sure the model is about the result.

The service handles 50 million requests a month through an API. An API lets other software send requests and receive results. Interfaze runs its frontend and backend on Vercel in the same codebase. The frontend is the part users see, while the backend handles requests. When a model writes code that needs to run, Vercel Sandbox runs it separately. Fluid compute helps the service scale as traffic changes.

Interfaze also uses models from outside providers to clean large amounts of data. A provider can stop responding or limit how many requests it accepts. AI Gateway can then send requests to another provider automatically. The team can also compare providers before choosing where to send work. Even providers offering the same model may return results of different quality.

The company uses the Workflow SDK for emails based on product use. The system checks which API a customer uses and at what volume. It compares that activity with the customer's first week of use. Workflows can retry failed steps instead of losing the whole sequence. They can also pause and continue months later while keeping their saved state.

Interfaze runs several model versions, with a separate URL for each one. Each deployment has its own logs, so the team can connect an error with the correct version. Preview URLs also let customers try a new version before release. The article describes several connected parts of operating an AI product. These include checkable results, provider changes, saved workflow state, and records of each release.