How Interfaze serves 50 million requests a month on Vercel
Interfaze, a provider of deterministic AI models for document processing and classification, migrated its entire stack to Vercel to handle 50 million monthly requests, including API routing, sandbox execution, and failover for third-party model providers.
Interfaze builds specialized models for tasks like OCR, document processing, and classification, where general-purpose LLMs often fall short. Founder Yoeven Khemlani notes that while LLMs excel at creative tasks, they lack the determinism required for fields like healthcare or legal work. Interfaze’s models return structured outputs—such as bounding boxes and confidence scores—alongside results, enabling traceability. The company’s backend, including the API handling 50 million monthly requests, runs entirely on Vercel, eliminating the need for separate infrastructure.
Interfaze’s backend and frontend share a single codebase on Vercel, with serverless functions processing all API requests and file uploads. Computation-heavy tasks, such as web scraping via Puppeteer, are offloaded to Vercel Sandbox, which executes code in isolation without additional setup. The platform’s scalability and high memory and timeout limits accommodate Interfaze’s workloads, including large-scale data cleaning using third-party open-source models. Vercel’s serverless architecture allows the system to scale automatically with traffic demands.
To mitigate risks from third-party model providers, Interfaze uses Vercel’s AI Gateway for automatic failover and provider switching. The gateway routes requests to alternative providers when one throttles or fails, ensuring continuous operation. It also enables A/B testing by allowing Interfaze to direct traffic to different providers and compare performance. This system replaces manual intervention and reduces downtime, improving reliability for customers in legal, healthcare, and finance sectors.
Interfaze leverages Vercel’s Workflow SDK for lifecycle management, including automated email systems triggered by product usage. The system tracks API calls, volumes, and compares them against initial usage patterns, with retries handled automatically. Vercel’s durable execution ensures workflows can pause and resume months later without losing state, a feature not consistently available in other systems. Additionally, each model version is deployed as a separate URL on Vercel, simplifying error tracking and enabling preview deployments for customers before full releases.