How Tavrn multiplies output across every function inside HIPAA constraints

Tavrn is building the retrieval layer for medical records in the US, connecting to providers nationwide — from the largest health systems to the long tail of clinics that still rely on fax, paper, and phone calls. Today, Tavrn applies that infrastructure to the legal industry, where its agents power the evidence layer for law firms: a firm hands over a case, and Tavrn's agents chase down records across thousands of providers and institutions, handling requests, follow-ups, and invoices along the way and delivering what the legal team needs to move the case.

Internally, Tavrn applies the same agentic model to how it runs the company. Eight engineers direct the work while a fleet of Devin agents executes in parallel across engineering, operations, customer support, and sales — dispatched from Slack, working through the night, and reviewed by humans. Because Tavrn handles protected health information, the entire system operates within its HIPAA requirements.

The result is leverage across the company: Tavrn's eight-person engineering team says it now ships like a team of fifty. The median time from a production alert firing to an agent investigating it is 35 seconds. And four out of five questions that ops, CS, and sales used to route to engineering no longer reach an engineer.

The problem: eight engineers carrying an operations company

Tavrn is unusually operations-heavy for its size. Alongside eight engineers are about ten people in operations who run pilots, record requests, invoices, and client touchpoints and handle sensitive client records under strict controls, plus roughly thirty people in sales.

Tavrn's engineers were already using coding agents in a local CLI. But local agents had a ceiling: they worked only while an engineer was at the keyboard, handled one task at a time, couldn't be shared as a link, and couldn't be opened to teammates outside engineering with access scoped and audited.

That left eight engineers carrying product development, bug fixes, and every alert the platform generated, while questions from the rest of the company about the code or data landed on them too.

Moving agents to the cloud could remove those constraints, but Tavrn holds protected health information. Any system touching its codebase or database had to pass its security and HIPAA review. The requirements were specific: a dedicated instance, SSO, audit logs, access scoped by team, agents running in the cloud, and the ability to reach them from Slack.

Tavrn evaluated the cloud-agent platforms on the market, and Devin passed those requirements. CTO Vitor Vavolizza set it up himself and started with his own work.

"Within two days I had a hundred agents working while I slept, and I realized the bottleneck had quietly moved. It was no longer how fast I could write code. It was how fast I could decide what should exist."

— Vitor Vavolizza, CTO, Tavrn

Each Devin session runs on its own cloud machine, with a shell, editor, browser, and running copy of Tavrn's platform. That lets a hundred sessions work at once while nobody is watching. Tavrn runs Devin in its own tenant with SSO and audit logs, with access scoped by org: engineering, operations, and sales each reach exactly what their role allows.

Vitor brought Devin to engineering next, and the whole team was using it by week four. Then operations. Then leadership. In its first six weeks, the fleet ran more than 1,600 sessions across the company.

The solution: an agent fleet behind every function

Devin now works across Tavrn's engineering, incident response, product improvement, operations, customer support, and sales workflows.

Each team can dispatch agents independently, with access scoped to its role. Engineers use Devin to build and validate features in parallel. Production alerts trigger investigations automatically. Operations uses Devin to answer questions and investigate issues before escalating them. Sales can call on the same system directly from Slack.

Engineering: Devin became the #1 contributor in its first six weeks

In its first six weeks, Devin opened more than 1,800 pull requests at Tavrn, with every one still passing human review and CI. Of the PRs the team has accepted or rejected so far, roughly 80% were merged.

That made Devin the company's top contributor by a wide margin: in the same window, Tavrn's most prolific human engineer merged 80 PRs. Weekly merged PRs increased from roughly 96 to 240, and in the most recent month, 93% of everything Tavrn merged originated from a Devin session. Vitor estimates the overall leverage at 5–10x because much of the gain comes from work that never lands on an engineer's plate in the first place.

Feature work accelerated too. Tavrn shipped new product features at twice its previous rate over the same six weeks, with agents carrying implementation while engineers stayed focused on design and review.

For complex features that cut across the platform, Devin runs Tavrn's product in a browser on its own machine with scoped test credentials, exercising the feature the way production would on top of the unit and integration suites. The work comes back validated end to end. Because every session has its own machine, several features can be built and validated in parallel. Tavrn now ships and iterates on three major product features a week.

"Our team's output is about 6x'ed since Devin. We're eight engineers, but we ship like fifty, and the reason is parallelism: at any moment our team has half a dozen features in flight, each with several sessions working in their own dev environment. Work that used to serialize through one engineer now just runs concurrently."

— Alex Pedersen, Head of Record Retrieval, Member of Technical Staff, Tavrn

The platform improves while the team sleeps

Every time Tavrn considers switching models in its own agents and pipelines, it runs an internal benchmark suite. That used to mean an engineer standing up a provider, running evaluations locally, and waiting. Now Devin runs the suite overnight, and the results are waiting in the morning, including the cost delta of making the switch.

A second automation runs daily against the product itself. Devin pulls the day's usage, finds cases that failed Tavrn's internal quality judge, and investigates how the system should have handled them. The core product improves continuously, with no engineer assigned to watch it.

Incident response: a 35-second median

Tavrn connected Devin to Grafana, Sentry, Temporal, and PostHog. Alerts land in a channel where Devin picks them up the moment they fire, reads the metrics and logs, determines whether something unusual is happening, and pages the team when it is. Devin runs the same triage on security alerts.

Because sessions keep working while the team sleeps, the on-call engineer often wakes up to a small PR with a proposed fix already waiting. A human reviews every fix, and high-severity incidents still page a person, arriving with a root cause and candidate fix attached.

Nearly nine in ten production alerts now receive a full in-thread investigation, and three quarters are under investigation within five minutes of firing. Before agents, an eight-person team couldn't deeply triage every alert as it arrived. The first hour of an incident belonged to whichever engineer could be pulled away from product work. Now investigation starts automatically, around the clock.

The same pattern covers bug reports, PostHog signals, and product suggestions. Anything larger becomes an engineer working with Devin to build the feature.

Every function, not just engineering

Four out of five questions that operations, customer support, and sales used to route to engineering no longer reach an engineer.

Instead, teams ask Devin directly. It reads the code and data their role is authorized to reach, with least-privilege access and full logging, to answer questions like: What's the live status of this pilot? How is this record request progressing? What does this specific case need next?

Tavrn's product itself is fully agentic, so operations can use Devin to understand the work of the agents running its retrieval platform. In an eight-engineer company, Tavrn sees the questions Devin absorbs as the equivalent of getting an engineer back. Teams without authorization cannot reach protected health information, and anything Devin can't answer or that touches PHI-sensitive systems still escalates to engineering.

"The most underrated part isn't the engineering gains. It's that our operations team now has a great engineer they can talk to at any hour. The distance between having a question and having an answer went to zero, and that changes how a whole company behaves."

— Vitor Vavolizza, CTO, Tavrn

Tavrn's head of operations became the second-heaviest Devin user in the company. Ops built its own automations on top of the access Vitor scoped for them: checking on pilots, pulling the live status of a record request, flagging customer support when a client needs a touchpoint, and identifying overdue invoices.

With a local CLI, giving non-engineers that reach would have meant handing out credentials and standing up a VPN, which a company holding PHI cannot do. With Devin, people log into their org and reach exactly what their role allows. When something genuinely needs engineering, ops investigates first and hands over the Devin session link, so the engineer starts from the investigation rather than a description of the problem.

That access has also changed what Tavrn buys. The company skipped the BI layer entirely because Devin answers data questions through the same scoped, audited access. Vitor trusts the answers because the agent understands the codebase behind the data.

Sales works the same way. When an account executive needed a product one-pager for a client, Tavrn's VP of Sales tagged Devin in the Slack thread and had the sample about a minute later. Engineering was never involved.

Results

Tavrn started with eight engineers carrying product development, incidents, and questions from across an operations-heavy company. Six weeks later, Devin is embedded across engineering, product improvement, incident response, operations, customer support, and sales.

The practical change is that far less work has to queue behind an engineer. Features can be built and validated in parallel, alerts are investigated automatically, and teams outside engineering can answer many of their own questions through scoped, audited access. That lets Tavrn keep a small engineering team focused on design, review, and product decisions while agents handle much of the execution underneath them.

"For most of software history, ambition was rationed by headcount. Devin is changing that equation for us. The interesting question now is how much a small team that really embraces agents can do, and I don't think anyone knows the ceiling yet."

— Vitor Vavolizza, CTO, Tavrn