Community Spotlight: Merlijn van Breugel & Iman Hashemi

How Ditto makes medical language easy to understand with AI

By Ioanna Lykiardopoulou

Every month, we spotlight one member for our community. At KickstartAI, we believe that meaningful progress starts with individuals — builders, researchers, founders, and practitioners — who are applying AI in bold, practical ways. Our Community Spotlight series celebrates their work, their challenges, and the lessons they’re learning along the way.

This month, we’re featuring Merlijn van Breugel, co-founder and CPO at Ditto, and Iman Hashemi, Head of AI & Data. They're on a mission to bridge the healthcare communication gap.

Visits to doctors or hospitals often come with an added challenge: making sense of the medical information that follows. This is exactly what Merlijn van Breugel, Iman Hashemi, and the Ditto team have set out to change. 

Founded in 2024, Ditto has developed a free app that uses AI to translate complex medical information into plain language. The aim is to help people better understand their health and diagnosis, regardless of medical condition, language differences, or age.

Users can record their doctor appointment to receive clear AI summaries, or scan medical documents to get clearer explanations. They can also share the information with family and friends. But how does it all work, especially in a field where accuracy and trust are paramount? 

Under the hood 

Ditto has opted not to build its own LLM from scratch, but instead add intelligence to foundation models.“We can go faster and differentiate more if we use the best model out there in the best possible way rather than spend all our resources on developing our own,” says Merlijn, the startup’s co-founder and CPO.

Big models are only one part of Ditto’s AI system, which also comprises Juvoly's Grutto model for speech recognition, a proprietary library for LLMOps, and various external tools for governance.

A core part of the system is explainability, adds Iman, Ditto’s Head of AI and Data. “For every AI summary, we need to know exactly what happens. What models we use, what versions we use, what decisions were made in the whole pipeline.” 

The second key element is the ability to develop or add new technologies very quickly. “We've created this template where if you want to add something new, you can very easily just add it on top of basically one file, run it, and get it to production fast,” Iman says. 

Evaluating the AI summaries 

To assess the quality of its AI summaries, Ditto has developed a framework of metrics with input from external experts. This framework combines two complementary methods: human evaluation through a panel of doctors and patients as well as automated assessments. 

Part of the automated analysis involves measuring how confident the model is in what it generates, based on the probabilities it assigns to each predicted word. This helps flag uncertainty in cases where the input is ambiguous — for instance, when a doctor mentions two different dosages. 

Ditto also uses LLM judges to evaluate overall quality, such as completeness and understandability, while more traditional NLP methods run similarity searches to verify details. “For example, to cross-check medicine names for accuracy,” Iman says.

The hallucination challenge

Among all metrics, hallucinations are the hardest to manage, Merlijn says. “They’re also the most serious aspect, and this is where most of our attention goes to.” 

One significant risk factor is incorrect medication details.“You can have a perfect summary but the wrong medicine name, and that can be very harmful,” he explains. For this reason, if Ditto’s system detects inconsistency in drug names or dosages, the app doesn’t generate a summary. “It’s annoying, but safer. I’d rather share no information than information I know is wrong.”

But not all hallucinations are clear-cut. Some are harmless, like when the model adds generic advice such as “call your doctor if symptoms get worse.” “Some doctors like that,” Merlijn says, “others don’t, because they didn’t explicitly say it.” 

“This is why we have multiple guardrails in place and a severity classification of hallucinations,” Iman adds. “We only want to share summaries that are safe and we prefer being on the conservative side.” 

Building trust 

Building trust into an AI product for healthcare is no easy task, but as Merlijn explains, Ditto benefits from being closely connected to the Dutch healthcare system. 

Through both professional and personal ties, the team collaborates with different organizations and academic hospitals not only to present its solution, but also to engage in open dialogue about how to build it responsibly. 

“Most of the doctors are very positive; others are more skeptical and have expressed privacy concerns about the recording feature,” Merlijn says. “What’s been working well for us in encouraging clinicians to test our app themselves. That has convinced many to start suggesting Ditto to their patients!”

Gaining user trust has to do a lot with clarity, he adds. “We try to be very clear on how to use the app, communicate why certain things wouldn’t work, and be transparent about privacy and data protection. We have consciously decided to not store medical data centrally, such that it is only stored on-device. Our users love that.” 

Launching an AI product in less than a year 

The startup launched its app in July this year, approximately seven months after the team started working on the product idea. 

“It all starts with having the best people,” Merlijn says, adding the importance of creating “maker time” for the development team. “This means having the ability to make quick decisions instead of involving multiple people in endless discussions — and trusting each other to make the right call.” 

Executing at speed is also about scoping things down and embracing ambiguity. 

“In the beginning, I struggled with that because I was used to roles with extensive  structure and plans. But being comfortable with ambiguity allows you to be more agile: if you suddenly have a better idea, you can directly change course and move with that idea.”

Ditto has set big ambitions for the future. “There’s plenty of room to grow in the Netherlands next year, and I’m quite certain that there will eventually be international expansion,” Merlijn says. Product-wise, the startup will also focus on summary personalization and on extending the functionality of sharing with loved ones. 

“This first year has definitely been a rollercoaster, with some hard lessons along the way,” he concludes. “But building a startup is the most fun I’ve ever had. Above all, working on something meaningful, something that helps people in such a tangible way, is what truly drives the team."

💡 Want to learn more about the KickstartAI community and our mission to accelerate AI adoption in the Netherlands? Check out our website, listen to our podcast, or follow us on LinkedIn!