EBANX: From one engineer's evaluation to 92% of merged PRs

The constraint on growth

EBANX processes payments across more than 20 markets in Latin America, Africa, and Asia, enabling international businesses to accept hundreds of local payment methods without establishing a local legal entity. The company's growth is driven by expansion into new markets, and every new market requires integrating dozens of local payment providers, each with its own API, protocol, and regulatory requirements.

Each integration took a team of engineers weeks to build and verify. There were more markets to enter than engineers to build them, and engineering capacity actively constrained how fast EBANX could expand.

The story of how EBANX broke that constraint starts with a single engineer and an unusual assignment.

An unusual assignment

When EBANX's VP of Engineering (now CTO) Kalecser Kurtz told Marlon Brandão de Sousa to stop writing code by hand, Marlon assumed he was being fired.

"I said, 'Okay, you're firing me.' He said, 'No, we have a tool called Devin that I want you to test. I want you to tell me clearly whether this is going to help us, and how.'"

— Marlon Brandão de Sousa

Marlon is trusted with evaluating AI tools and helping shape engineering strategy at EBANX, a responsibility earned over a two-decade career building complex systems for Brazil's largest television network, Itaú's internet banking platform, and EBANX's global payments infrastructure.

He has also been blind since birth, which makes him an unusually demanding judge of how autonomous an AI engineering tool actually is. Where a sighted developer sees a red squiggle for every syntax error, Marlon faces a multi-step sequence: build, wait for compiler output, listen to the screen reader describe errors, go to the offending line, apply the fix, rebuild, and verify. The AI assistants he had previously tested produced error-ridden code, which ballooned his time debugging.

"Even though you have sound assistance, you see no colors. There is nothing helping you concentrate on the things you should be concentrating on. I needed to push my brain to its limits to survive in the arena."

— Marlon Brandão de Sousa

Devin was the first tool that verified its own work end-to-end. Each Devin session runs on a dedicated cloud machine with a shell, an editor, a browser, and a build environment, where it iterates autonomously until its work is merge-ready and returns a finished PR with a full session trace.

"Devin will code, test the code, get the errors, fix the errors, test again, make sure the code works, apply linters, and open a PR. No more manual switching between this and that for me."

— Marlon Brandão de Sousa

Marlon spent a month running Devin on real tasks and studying the full execution logs of every session: how it reasoned about code, recovered from errors, and handled ambiguity.

"I read these logs very carefully because I was really trying to understand how the system worked. By the end of that month, I was quite confident that a revolution was going to happen."

— Marlon Brandão de Sousa

The first production project

To prove Devin's value to the rest of the organization, Marlon chose a project critical to the business: a payment provider integration, exactly the kind of work that was holding back EBANX's expansion. Each new provider must implement a set of typed internal interfaces, communicate with the provider's external API, and include comprehensive unit tests.

He approached the senior developer leading the project privately with an offer:

"If this works, the project will be ready in 20% of the estimated timeline. Everyone will ask how you did it. And if it fails, you've lost one day, and you still have the remaining weeks to deliver normally."

— Marlon Brandão de Sousa

The team wrote specifications for the internal interfaces, then pointed Devin at three inputs: the provider's API documentation, the interfaces to implement, and EBANX's internal reference examples.

"We went for a coffee. When we came back, Devin had carefully reviewed the provider API documentation and proposed a complete plan: which interface maps to which endpoint, mocked responses from the documentation, unit tests, everything."

— Marlon Brandão de Sousa

The project, scoped for six weeks, was delivered by Devin in about twelve hours. The senior developer was convinced enough that he presented the results to the organization himself.

Since then, EBANX's SDK-based integration framework, built with Devin, has compressed typical provider integration delivery timelines. With EBANX continuously onboarding new payment providers across Latin America and beyond, these time savings compound across dozens of integrations per year.

"We stopped asking 'how can Devin help me code faster?' and started asking 'how do we structure our code so Devin can actually deliver more?' That mindset shift changed everything. And now, when someone shows me a great result, I don't ask for the code anymore. I ask for the prompt."

— Bruno Rossetto, Director of Product Engineering, EBANX

From one project to the default way of shipping

After that first project, adoption spread organically among junior and mid-level engineers. Senior engineers were initially skeptical, but they were won over by how Devin tested its own work, similar to a human QA engineer — launching the app in a browser, clicking through the affected user flows, and taking screenshots to verify UI elements.

Devin grew to support EBANX's most revenue-critical systems. The highest-volume use case became API modernization across the core payment services. For large framework migrations, teams launch a Devin session for each module, component, or repository and run the whole migration in parallel. Devin's automations let it work continuously in the background: it drafts a post-deploy monitoring plan on every approved PR in the core payments repo, runs periodic code audits, and triages production alarms, posting its diagnosis back for the on-call engineer.

Devin is now a default choice at EBANX. In the core payments platform, where most of EBANX's engineering work lands, 92% of merged PRs are created with Devin.

EBANX even instituted a "nudge system," an internal review process that flags every pull request not authored by Devin and asks the engineer to document why they chose to write the code manually.

Building trust in a highly regulated business

As a payments company, signed commits, role-based access control, and audit trails are non-negotiable.

Where troubleshooting once meant tracking down and interrogating the original code author, engineers across the org can now open a Devin session and immediately read the thinking behind each line. When the CTO Kalecser Kurtz wants to understand a change, he opens the Devin session behind it and reads about how the work was done.

"Session preservation is the most precious artifact."

— Kalecser Kurtz, CTO, EBANX

Devin Review adds another layer of confidence, automatically reviewing the code before engineers see the finished PR. Before each deployment, engineers consult Devin Review's insights as guidance on what to watch in production once the change goes live.

"Its accuracy is unbelievable — it's almost witchcraft. Every bug it identifies tends to be a real bug that was avoided downstream, where it would be more expensive to fix. It is the best review agent I have ever used."

— Marlon Brandão de Sousa

EBANX also invested in the guardrails that keep this volume of work safe: integration tests, type safety, and linter enforcement.

Beyond the pull request

Beyond coding and review, Devin is now impacting how EBANX designs and specifies its software. Product owners and engineers review design docs and building specifications together in multiplayer Devin sessions, where Devin surfaces technical constraints in the chat.

For Marlon, the change is personal. Cloud execution lets him launch several tasks in parallel, freeing his time to mentor junior developers and strategize with product teams. When he returns, the pull requests are ready to be approved and merged. Everything Devin produces — its planning, reasoning, the commands it ran, and the test results — comes back as plain text, which is exactly how screen readers process information, so Marlon has the same complete visibility into Devin's work as any sighted engineer. Because Devin works over tools like Slack, he can put it to work from anywhere:

"I am extremely productive at night. I have an idea, open Slack, start a new Devin task, pass some prompts, go to sleep — and the following day I have something I can evaluate."

— Marlon Brandão de Sousa

Twenty years ago, a piece of software gave Marlon access to a computer for the first time. Today, he is shaping how an entire global organization builds software with AI, in an operating model where an engineer's thinking, judgement, and cross-functional collaboration matter more than ever.

_{*Estimate based on Cognition's internal metrics}