Delphi rebuilt its Python backend on Vercel and now ships 100+ deploys a day

A case study published on the Vercel blog describes how Delphi, a startup that converts a person's writing, recordings and teaching into an interactive “digital mind”, now pushes more than 100 production deployments a day with ten engineers and no dedicated infrastructure role. The company rebuilt its Python backend on Vercel roughly six months before publication, having run its frontend there since launch.

The backend was the bottleneck

Delphi's backend previously ran on AWS using ECS, Docker Desktop and local databases. That setup was tolerable while the team was tiny and context lived in people's heads, but onboarding a new engineer required a full day of environment setup before a first deploy. Founding engineer Spencer Schoeben says the team could have fixed the problem without leaving AWS, but a proper fix would have meant designing and owning the infrastructure itself, which was not the company's goal. After the migration, new engineers reach production through the same everyday deploy workflow the rest of the team uses.

Workflows and Queues made the move possible

Two Vercel features that did not exist when Delphi was founded made the migration feasible, according to the post: Workflows, which handles long-running work, and Queues, which handles background jobs. The product depends on both — digital minds need agents that prepare content in advance, plus ingestion pipelines that convert a person's books and recorded talks into a knowledge graph. Developers write a workflow or queue handler as an ordinary function in the codebase, and Vercel provisions whatever is needed to run it at deploy time. “With Vercel, we barely think about infrastructure at all,” Schoeben told the Vercel blog.

Straight to production behind feature flags

With backend and frontend on one platform, Delphi dropped its staging environment entirely. Changes now go to production more than 100 times a day behind feature flags, with A/B tests replacing batched releases. Vercel Agent's anomaly detection watches production at that pace, flagging misbehaviour and pointing to a cause — problems the team says it often would not have caught before. The faster cadence has cleared a backlog of small experiments that never made the priority list, many of them testing how to convert a visitor into someone who starts their own mind. Shipping is not limited to engineers, either: the CPO and growth teams deploy dashboards, prototypes and experiments themselves, which Schoeben says would not have happened on the old setup.

Agents as the primary platform users

Very little development happens locally. Engineers hand problems to cloud agents and first see the result in a preview deployment, each with its own live URL that can be checked from a phone or a Slack thread. Agent failures usually trace back to missing information, so Vercel exposes logs, deployments and environment variables through its SDK, MCP and CLI — and Schoeben says agents are now the main users of those interfaces. Delphi also runs an internal support agent on Vercel Sandbox, built with eve, Vercel's agent framework. Customer success asks it in Slack about reported issues and receives a root cause and a proposed fix; an earlier version on a hosted agent platform was limited to that platform's built-in integrations, whereas in eve the tools are simply code Delphi writes. Chat traffic passes through AI Gateway, letting the company choose a model per digital mind, adjust routing as its own evaluations change, adopt new models the same day they ship, and rely on the gateway for failover instead of arranging fallback capacity directly with providers.

Delphi's next step is an interface beyond chat: a search page that answers a question with perspectives from several minds at once and adapts to whoever is asking, running as long-running agents on Workflows.

Why it matters

This is a vendor case study, so the framing favours Vercel and independent verification is limited. Even so, the operational details sketch a delivery model that is becoming common: staging replaced by flags and previews, non-engineers deploying their own work, and platform APIs designed for AI agents as much as for humans. It is also a concrete data point in the debate over whether serverless platforms can host serious backend workloads. Delphi's experience suggests the answer hinges less on plain functions than on newer primitives — durable workflows, queues, sandboxes — which indicates where the practical bar for running a real backend on serverless infrastructure now sits.