I Sold My Startup A Year After Founding It. Here’s Why That Was The Fastest Way To Build Real-World Healthcare AI

By Louis Blankemeier

In October 2024, my co-founders and I set out to make our Ph.D. research useful in the real world. We had built AI models that could interpret medical images such as X-rays and CT scans across tens of thousands of potential diagnoses, generating comprehensive radiology reports that mirror how radiologists reason in clinical practice. At a time when AI in radiology was limited to flagging a handful of specific conditions, this marked a fundamental shift.

Less than a year later, we faced a critical fork in the road: raise venture capital and continue independently, or accept an acquisition offer from Radiology Partners, the world’s largest radiology practice.

The conventional wisdom in tech is that real ambition means staying independent. But in asking ourselves what it would truly take to transform healthcare, the answer was different.

Clinical AI is highly regulated with long sales cycles and complex stakeholder dynamics, where structural advantages tend to harden market positions and compound over time. We decided that joining forces — carefully structured to protect our velocity — would dramatically improve the odds that we realize our mission of significantly increasing the world’s access to healthcare.

Research success is not clinical readiness

During my Ph.D., I trained radiology AI foundation models on what, at the time, felt like massive research-scale datasets; tens to hundreds of thousands of studies. These models make for strong academic demonstrations, prototyping new capabilities across a range of tasks. In real clinical settings, however, they would not yet have met the standards required for production-level safety and consistency in patient care.

Despite the persistent narrative that AI will make radiology obsolete, the reality is that the problem is extraordinarily difficult. A single CT study, for example, can contain 10 high-resolution volumetric series, effectively 3D videos. Add prior studies for the same patient, and you can have a billion pixels of data.

Those billion pixels encode entire medical textbooks worth of information. On top of this, real-world radiology is defined by edge cases where rare but critical pathologies are encountered regularly. We learned a hard truth early on: Models that work in controlled research environments often fall apart when exposed to real-world complexity.

Think about self-driving cars. A decade ago, progress looked impressive. But the real world kept introducing new failure modes. After more than a decade of significant capital investment, only a handful of companies have approached true reliability.

Components required to build reliable models

Key patterns emerged. The companies that made the most progress controlled the entire system and achieved scale early. They owned the vehicles, the sensor stack, the data collection pipeline, the simulation environments, and the deployment infrastructure. That integration, operating at scale, allowed them to continuously collect rare edge cases, retrain models, validate improvements and redeploy safely.

Radiology is no different. Success in the real world requires massive, diverse historical datasets and live data feeds that continuously surface rare edge cases and distributional shifts. It requires vast clinical resources and operational infrastructure to redesign clinical workflows around AI, engineer systems that perform reliably at scale, conduct large-scale research studies, secure regulatory clearance, refine models safely, and continuously monitor performance post-deployment.

Additionally, frontier language models have clearly demonstrated that continuous, high-quality and extensive human feedback is the secret sauce in making models useful. This is no different in radiology. In a world where radiology reports are drafted by AI, every draft must be reviewed, edited and signed off by a human radiologist.

Those edits become high-quality signals that can be leveraged for improving the AI models. Better models elevate radiologists’ accuracy and capacity. Improved radiologist accuracy increases the quality of future training data. Increased capacity allows radiologists to take on additional contracts.

That, in turn, generates more data and high-quality corrections, setting a powerful flywheel in motion. Access to this correction data is rare in AI and can only work meaningfully at a massive scale. These capabilities would be incredibly difficult to achieve as a standalone AI startup.

In healthcare, growth follows evidence

In healthcare, trust is hard earned. It rests on demonstrated clinical efficacy, reliability, security and regulatory rigor. For a health system or radiology group to adopt technology from a new startup, particularly in workflows that directly affect patient care, requires rigorous, real-world evidence.

Evidence in healthcare is not generated in small pilots. It is built through sustained performance across diverse sites, patient populations, modalities and edge cases. If a system proves itself within the world’s largest radiology practice, it establishes credibility across multiple dimensions at once — efficacy, reliability, security and scalability.

In sectors where lives are at stake and the goal is to build something that endures, the way to build it is from within the system you’re trying to improve. Selling early didn’t shorten our journey, it accelerated it. It gave us the foundation required to deliver on our mission of significantly increasing the world’s access to healthcare.

Louis Blankemeier is the CEO and co-founder of Cognita, the AI business unit of Mosaic Clinical Technologies at Radiology Partners. During his undergraduate studies in physics and electrical engineering, he became driven by a singular mission: increasing the world’s access to healthcare through technology. Convinced that AI was the most promising technology to make this happen, but not yet good enough for real-world clinical use, he pursued a Ph.D. in AI at Stanford University where he focused on foundation models for radiology. His doctoral work produced Merlin, a 3D vision-language model for CT interpretation published in “Nature” in 2026 and recognized as one of the most important papers in the field.

Illustration: Dom Guzman