If I say “Research and Development”, probably you’d think of a lab with people wearing goggles, using microscopes, and… well, you get the point. That picture disappears the moment you discover what an R&D department does at an AI consulting firm, or, let’s be honest, at Arionkoder. 

The truth is that R&D has everything to do with AI. And here’s proof from our own R&D team. They found that when a hospital or research team has a set of labeled medical images, getting to a first model that performs well enough to be taken seriously takes a specialist in that exact area about 21 days (for those who aren’t familiar with this, a model in computer vision for healthcare can even help detect diseases faster and assist doctors). The other issue, beyond this task being time-consuming, is that the specialists who work on this are scarce. 

With that in mind, our R&D team built a harness for an AI coding agent that does the same job in 2 to 7 days, with a person reviewing the work instead of running it fully. 

The paper describing this project was accepted to a workshop at MICCAI 2026, the leading international forum for medical image computing and computer-assisted intervention. It’s the first paper from our R&D department to be accepted at a conference, and the second manuscript produced by Arionkoder.

The two authors behind this project, José Ignacio Orlando, R&D Director, and Eugenia Moris, Data Scientist, sat down with us to discuss what they built and what they were most surprised about.

What’s the "AI Scientist" workflow, and how does it build a medical imaging model?

The “AI scientist” workflow is an agentic pipeline, orchestrated through Claude Code, that takes a raw labeled dataset and returns a trained, competitive baseline model with light human oversight. “During my PhD, I’ve worked extensively with medical imaging, so I know firsthand how much time and effort it takes to develop a computer vision model. It’s not just about solving the task itself, you also need to understand the medical context, what clinicians are looking for, and make sure the model can generalize well across different data sources”, discusses Eugenia Moris, first author of the paper, about how the project started.

The main reason for this is the bottleneck anyone working in medical imaging recognizes. “We identified a recurring problem in computer vision and especially in medical imaging: every time a new dataset or clinical problem appears, you often need a specialist in that specific area to build a strong baseline model,” explains José Ignacio Orlando, the corresponding author of the paper. “The challenge is that computer vision talent is scarce, and the more specialized the medical problem becomes, the harder it is to find someone with exactly the right expertise”, adds.

A strong baseline has to do more than run perfectly. It has to perform well enough to become a serious starting point for evaluation in a real clinical setting, whether the task is classification, segmentation, or detection. Before this project, reaching a first model candidate took the team around 21 days of expert work.

The workflow combines literature review, automated code generation, and evidence-driven experimentation. How does it look in practice? The AI agent follows the same process an experienced computer vision engineer would: understand the problem, research the relevant literature, identify suitable architectures, choose which experiments and hyperparameters are worth exploring, and then orchestrate training and evaluation. That includes the decisions that consume the most time, like tuning the learning rate, selecting data augmentation strategies, and comparing approaches systematically. 

“Initially, we focused on using AI agents to automate data augmentation, making that process more dynamic and efficient”, says Eugenia. “But as the project evolved, we realized we could go so much further. That ultimately led us to develop the harness to make the agent capable of generating a strong baseline for an entirely new medical imaging challenge with minimal human intervention”, points out Eugenia about the process. “We also built guardrails into the repository so that the agent consistently follows that well-defined process, which is ours, something that we’ve adjusted and perfected over the past 8 years,” finishes José Ignacio Orlando. 

About the human side of this process, our R&D Director underlines: “We added an explainability layer on top of the system. One part looks like a tree where the computer vision engineer can audit the sequence of decisions made by the agent, the experiments it considered, and the paths it followed. Another visualization shows how performance evolved throughout the process, including which experiments improved the model and which ones failed. This makes it much easier for a human to understand what happened and intervene when necessary.”

Eugenia Moris and José Ignacio Orlando, authors of the project.

What the agent achieved, and what surprised even the team

The workflow produced competitive baselines for four public medical imaging tasks in 2 to 7 days each, down from about 21, and generalized across segmentation, classification, and detection. That time also includes the time spent training the model candidates on a single GPU, sequentially. Here are some numbers that add to that surprising achievement:

• Competitive baselines across four public medical imaging tasks, in 2 to 7 days each, down from 21 days.

• 6th of 15 teams on both PUMA tracks. PUMA is an open challenge on panoptic segmentation of tissues and cell nuclei in advanced melanoma images; its two tracks cover panoptic segmentation with three and with ten instance classes.

• 52 minutes to a first model candidate, down from about a week.

• Validation Dice from 32% to 68% on PUMA, in two days of agent-planned experimentation.

But what surprised the team the most wasn't the speed. It was that the agent made decisions they assumed required visual judgment. To test that, they ran a blind head-to-head on data augmentation, the practice of modifying training images through brightness, rotation, or contrast changes, where the risk is producing artificial images that hurt the model or misspend part of its capacity. Eugenia independently explored augmentation strategies and selected the combination she considered best. The agent was then asked to solve the same problem without seeing her choice.

“Seeing a coding agent independently select an augmentation strategy that outperformed a very experienced human on the primary performance metrics was definitely one of the most surprising results of the project,” says José Ignacio Orlando. 

“What also surprised me was how well the agent performed once we gave it the right foundation. By providing a clear framework, the right context, and well-defined skills, it was able to make many of the same decisions that I would have considered myself”, adds Eugenia.

Humans aren’t left outside the loop. “Like any AI system, agents can occasionally hallucinate or make questionable decisions, so having someone review the overall direction helps ensure the project stays on the right track,” explains Eugenia.

Scarcity plays a different role here because the nature of the specialist’s work is different. “Instead of being deeply involved in one project at a time, the same person could potentially orchestrate five different projects in parallel while the agents independently explore and run experiments”, discussed José Ignacio Orlando. “I think that’s one of the most important contributions of this approach: we can now have a general ML practitioner auditing these decisions, asking for more context if needed, and approving or correcting the agent without having to be an expert in that particular field. This is textbook upskilling, but applied in a very tough domain.”

Plus, this AI Scientist workflow goes beyond the walls of Arionkoder. "Whether you're working with skin images, retinal scans, histopathology, or another imaging modality, our goal is to have a reliable baseline ready in about a week," says Eugenia Moris. "What's especially important is that this baseline has already shown good domain generalization, making it a solid starting point rather than just a proof of concept. Once you have that baseline, researchers can spend their time on what really matters: testing new ideas, developing novel methods, and pushing performance further instead of spending weeks building the initial solution."

“This is about democratizing access to advanced computer vision”, says our R&D Director. “It can reduce both the cost and the time required to get to a first solid solution, and allow hospitals and research groups to test ideas that might otherwise be too expensive or too slow to pursue.” Cost and time are the same limits we design around in healthcare delivery, from an AI triage agent that cut ticket handling by 75% to a preventive care app taken from discovery to live product in seven weeks.

The workflow is now a reusable platform component that we use to better serve our Computer Vision customers, with 3D support for CT and MRI, and more coming next on the roadmap.

What this means for R&D at Arionkoder 

At Arionkoder, R&D expands what is technically possible by exploring frontier capabilities and turning them into reliable building blocks that delivery teams can reuse. It’s one of the three layers of The Foundry, our operating model, alongside Delivery and Innovation.

The "AI Scientist" workflow shows the full R&D cycle in one place. It began as a question about data augmentation calibration, grew into an agent that builds competitive baselines in days, earned external scientific validation, and ended as a reusable component. 

The scientific validation proves we’re on the right path. “This is the first external scientific recognition of this project," says José Ignacio Orlando. "It shows that the problems we're choosing to solve are relevant and aligned with where AI, and especially medical AI, is heading," continues Eugenia Moris. 

“What it does give us is more confidence to keep investing in this kind of work. We can build technology internally, evaluate it rigorously, publish what we learn, and then bring those ideas back into the way we develop AI systems for real-world problems,” concludes José Ignacio Orlando, probably already thinking about which wall to break next.