Thirty hours searching for one incident

Thomas Alflen, co-founder of Oddity.ai, heard that story and immediately recognised the problem. His Utrecht-based startup had spent years working on violence detection for Dutch municipalities and prisons. The question was whether that same technology would work in the American healthcare sector, where privacy regulations are stricter and trust in technology vendors is in short supply.

The answer turned out to lie not only in technology. It lay in being present: visiting locations, listening to staff and calibrating the AI to each environment.

From tracking influencers to a forensic psychiatric clinic

Thomas Alflen, Gerwin van der Lugt and Nick Mulder met during their studies at Utrecht University. All three in their early twenties, all three with ambitions to start a company. Their first idea was an AI tool to monitor the social media accounts of influencers [1].

"A fun toy, but we weren't happy with the impact it had on society", says Alflen. They went looking for a real problem.

They found it with the police. An officer sought help identifying incidents of violence in camera footage. The founders immersed themselves in darkened surveillance rooms and saw what was happening there. People trying to monitor too many screens at once. They decided to build an "untiring observer".

A coincidental confluence. Their first office was located in the former Pieter Baan Centrum in Utrecht, the Netherlands' most notorious forensic psychiatric clinic, where high-profile criminals awaited trial [1]. The algorithm trained within those walls has since detected tens of thousands of violent incidents.

Not catching the perpetrator, but helping the victim

What sets Oddity apart from most surveillance technology lies in a subtle shift.

"We develop our software primarily for the victim, who needs to receive help as quickly as possible. Municipalities sometimes tend to focus on catching the perpetrator.", Gerwin van der Lugt, co-founder [2]

The algorithm does not recognise faces or identities. It analyses body language, movement patterns and physical interactions to detect aggression. The AI sends an alert; a human decides. No personal data leaves the system; no footage is shared with third parties. Everything runs in an encrypted private cloud environment, fully HIPAA-compliant for the American market.

That architecture addresses a deeper problem than compliance alone. Healthcare institutions have too often been disappointed by vendors who promised security while quietly harvesting data for model training. When Alflen says "the data stays yours", that is not marketing.

In American healthcare, these figures are critical. Healthcare workers make up 13% of the workforce, yet account for 73% of all workplace violence-related injuries [3]. A 2024 study found that 81.6% of nurses experience at least one violent incident per year [4]. A staggering 88% of those incidents are never reported [5]. The problem is larger than anyone can see.

Oddity now focuses on care providers supporting people with disabilities. According to the Bureau of Justice Statistics, people with disabilities face a violent victimisation rate of 46.2 per 1,000, nearly four times higher than for people without disabilities. This risk is greatest among people with cognitive disabilities, who have the highest rate of violence: 83.3 per 1,000.

Crossing the ocean with a cookie

At the start of 2025, Oddity crossed the Atlantic. Not with venture capital and a sales team of fifty, but with a small team and a box of stroopwafels at every meeting.

For the uninitiated: a stroopwafel is a beloved Dutch treat, two thin, crispy waffle layers with caramel syrup in between. Dutch people bring them along everywhere as a gift. They have become a kind of national calling card.

"There are more opportunities there than in Europe", says Alflen [2]. But the approach is far from Silicon Valley. While others attempt to scale remotely, Oddity flies to the location: walking the corridors, meeting staff and calibrating the AI to each setting, so that the implementation fits the care provider, not the other way around. Ensuring a genuine operational fit between care providers and AI.

Their first American client, Amego Inc. in Massachusetts, had a concrete frustration. Amego had cameras, but no system for extracting actionable information from the footage in real time. They wanted a system that proactively alerts when significant incidents occur.

"Oddity's consultative approach stood out, they took the time to understand our needs and adapt to our environment", said John Randall, CEO of Amego Inc. [6].

The integration was plug-and-play. But the real difference came in what followed. Continuous communication, rapid adjustments, fine-tuning of detection thresholds. Too sensitive and staff start ignoring the constant false alerts. Too conservative and real incidents slip through. That calibration is more complex in behavioural care, where clients' movements can appear aggressive without being violent.

Those thirty hours of searching? A thing of the past. Incidents that previously went undetected for weeks now surface in real time.

Bootstrapped as a strategy

Still bootstrapped and deliberately measured in pace, Oddity used a Massachusetts-focused rollout last year to confirm product-solution fit with care providers. With that foundation in place, their focus has expanded to the United States as a whole.

In a sector where trust is fundamental, starting slowly and in person becomes a competitive advantage.

The AI market in healthcare is projected to grow from $29 billion in 2024 to more than $500 billion by 2032 [7]. Nearly all attention goes to diagnostics and drug development. Aggression detection remains a niche. And that is precisely where Oddity wants to be.

The lesson

The lesson from Oddity is not blitzscaling, but fit-first scaling. Prove it in the field, earn the trust of care providers, then expand. The company won early adoption by being present: visiting locations, aligning the system with real-world routines and supporting staff after go-live.

With live production deployments and results in hand, the question now is execution at national scale: can they maintain the same hands-on, provider-first rhythm as they grow across the US?

The edge here is not the loudest model or the largest funding round; it is reducing a 30-hour task to half a second, and then still picking up the phone to ask whether it really helped.