Delphi Deploys Python Backend 100 Times a Day with Vercel Workflows and Queues

Digital mind network startup Delphi is performing over 100 production deployments a day after migrating its Python backend from AWS to Vercel. Ten engineers participate in deployments alongside product and design team members without a dedicated infrastructure role, and the onboarding time for new engineers has been significantly reduced.

Changes in the Deployment Process

Previously, using ECS and Docker Desktop took a full day just for environment setup, but the switch to Vercel eliminated the burden of infrastructure ownership. The staging step is skipped, and deployments go directly to production via Feature flags, with real-time monitoring handled by the anomaly detection features of Vercel Agent.

Core Technology: Workflows and Queues

The key to enabling the backend migration was Vercel Workflows and Queues. Written as functions within the codebase, they are automatically provisioned upon deployment, requiring no configuration.

• Workflows: Handle long-running tasks such as pre-generating content
• Queues: Handle background jobs such as ingestion for knowledge graph construction

AI Agent-Friendly Infrastructure

AI agents can directly control deployments through Preview deployments, which provide a unique URL for each push, along with SDKs, MCP, and CLI. Vercel Sandbox is used to allow internal agents to analyze the codebase and, when the customer success team reports issues via Slack, generate root cause analysis and fix suggestions.

Model Management via AI Gateway

Through AI Gateway, models are dynamically routed based on the characteristics and self-evaluation results of each digital mind. Routing is possible on the day a new model is released, and automated failover features ensure fallback capacity in the event of a model provider outage. This improves time efficiency by addressing edge cases that were previously missed.