Delphi's platform problem: infrastructure the team refused to own
Delphi builds digital minds — capturing what someone has written, recorded, and taught so anyone can query that expertise on demand. Founding engineer Spencer Schoeben calls it "a destination where you can find and interact with interesting minds."
No such network existed before, so the company learns what to build by shipping features and watching usage. That approach only survives if the product changes quickly: "If we are going to win, we need to have an excellent developer and agentic experience," Spencer says. "Otherwise, we could not adapt to how fast our user's demands change."
The team is ten engineers and has no dedicated infrastructure role; everyone ships code, including product and design. The frontend ran on Vercel from day one — "the frontend has always been easy," Spencer says, "but the backend was a nightmare." Six months ago Delphi rebuilt that backend on Vercel too, and the whole team now pushes 100+ production deploys a day.
From AWS and a day of setup to a shared deploy workflow
The old backend lived on AWS on top of ECS, Docker Desktop, and local databases. It worked while the team was small and most context sat in people's heads, but it made onboarding slow: a new engineer needed a full day of environment setup before their first deploy.
The problem was fixable without leaving AWS, but doing it properly meant designing and owning the infrastructure. That was off the table. "That is not our current goal as a company," Spencer says. On Vercel, new engineers reach production much sooner, using the same deploy workflow the rest of the team runs every day.
Workflows and Queues: the primitives that made the move viable
Neither Vercel Workflows nor Vercel Queues existed when Delphi was founded. Without them, Spencer says, the migration would have been much harder, if possible at all — and the backend depends on exactly what they provide.
Digital minds need durable agents generating content ahead of time, and ingestion that converts a person's books and talks into a knowledge graph. Workflows takes the long-running work; Queues takes the background jobs. Neither needs setup, because a developer writes the workflow or queue handler as a function in the codebase and Vercel provisions what's required on deploy.
"With Vercel, we barely think about infrastructure at all."
— Spencer Schoeben Founding Engineer @ Delphi
Staging is gone; anomaly detection watches production instead
Once backend changes stopped looking like heavyweight releases, they became ordinary product work. Delphi previously deployed to staging and then production; it now skips the middle step entirely, running A/B tests instead of batching changes.
Vercel Agent's anomaly detection monitors production at that cadence, flagging misbehavior and its cause. Before, Spencer says, they often would not have known.
The faster loop cleared a backlog of small experiments that never reached the priority list — many of them testing how a visitor becomes an owner, arriving to talk to a mind and leaving having started one. Shipping isn't confined to engineers either: Delphi's CPO and growth teams build and deploy dashboards, prototypes, and experiments themselves, which Spencer says wouldn't have happened on the old setup.
Agents as first-class users
Little development happens locally at Delphi. Engineers hand problems to cloud agents, and the preview deployment — every push gets its own live URL — is where they first see the result, checkable from a phone or a Slack thread without pulling the branch.
"We couldn't keep up with AI on our old infrastructure. If it took a human eight hours to figure out, an agent would have given up."
When an agent fails it is usually starved of information. The SDK, MCP, and CLI expose the data and controls it needs — read logs, inspect a deployment, change an environment variable — through whichever interface it happens to be using. Agents are now the primary users of those tools at Delphi, so the platform has to be as approachable for them as for people.
Delphi runs its own agents on Vercel as well. An internal agent on Vercel Sandbox helps the customer success team work with the company's codebase: they ask it about a reported issue in Slack, and it returns a root cause and a proposed fix for an engineer. Edge cases that fell through the cracks before are now covered.
Sandbox supplies an isolated environment with a file system, so the agent can check out the codebase, run it, and pull in whatever data a question requires. The agent is built with eve, Vercel's agent framework. Its first version ran on a hosted agent platform limited to that platform's integrations; in eve, tools are code Delphi writes, so a missing capability is simply added.
Model routing and failover through AI Gateway
Chat traffic goes through AI Gateway. Every digital mind differs, and some perform better on one model than another, so Delphi selects the model per customer and adjusts as its own evals shift. A new model — open-source or not — can be routed to the same day.
Gateway also covers failover. Previously Delphi worked directly with a model provider to arrange fallback capacity on other clouds, which was hard. "With AI Gateway, we always know that we're going to get the fallbacks if they exist," Spencer says.
Beyond chat, on the same stack
Delphi's first chapter was training digital minds and letting people chat with them; it is now building the interface past chat. Spencer describes a search page where you bring a question, receive perspectives from several minds at once, and the page adapts to who you are. That work runs as long-running agents on Workflows, with AI Gateway supplying whichever model each experiment calls for.
"Everything we do is about moving really fast and punching above our weight. Vercel is a big part of making that possible."
About Delphi: Delphi is a platform that turns a person's knowledge, thinking, and voice into a digital mind others can talk to. Anyone can create their own Delphi to make their expertise discoverable and accessible, part of a network of minds people can seek out, learn from, and converse with.
