Less than a year into our startup’s existence, a critical juncture emerged, presenting us with two distinctly different paths forward. We could either pursue a significant venture capital funding round to maintain our independence and continue building our vision autonomously, or we could accept a compelling acquisition offer from Radiology Partners, a titan in the industry and unequivocally the world’s largest radiology practice. The decision weighed heavily on us, challenging the prevailing wisdom deeply ingrained in the tech startup ecosystem: that true ambition and transformative potential are inherently tied to staying independent, to building a standalone unicorn. However, as we rigorously interrogated what it would genuinely take to effect profound and lasting change in healthcare, the answer that emerged was surprisingly contrarian, leading us away from the conventional path.

The landscape of clinical AI is notoriously complex, characterized by stringent regulatory hurdles, protracted sales cycles, and intricate stakeholder dynamics. In such an environment, structural advantages are not merely beneficial; they tend to harden market positions, creating formidable barriers to entry and compounding over time for established players. After careful deliberation, we concluded that joining forces with Radiology Partners – an integration meticulously structured to safeguard our agility and developmental velocity – would dramatically enhance our probability of realizing our overarching mission: to significantly expand global access to high-quality healthcare through technology. This wasn’t merely a business transaction; it was a strategic alignment aimed at accelerating our impact.

Research Success Is Not Clinical Readiness: Bridging the Chasm

My Ph.D. journey involved training radiology AI foundation models on what, at the time, felt like monumentally massive research-scale datasets, often comprising tens to hundreds of thousands of medical studies. These models proved incredibly effective for academic demonstrations, showcasing novel capabilities and prototyping new functionalities across a wide array of tasks. They were excellent for proving concepts and pushing the boundaries of what was theoretically possible. Yet, the stark reality was that these same models, when transplanted into real-world clinical settings, would not yet meet the exacting standards required for production-level safety, consistency, and reliability in patient care. The leap from a controlled academic environment to the unpredictable, high-stakes clinical frontier is a chasm far wider than many realize.

Despite the persistent, almost sensationalist narrative that AI is on the verge of rendering radiology obsolete, the truth is that the problem of accurate, reliable AI-driven medical image interpretation is extraordinarily difficult. Consider the sheer data density: a single CT study, for instance, can encompass 10 high-resolution volumetric series, which are essentially 3D videos of the patient’s internal anatomy. When you factor in prior studies for the same patient, meticulously tracked for comparative analysis, you are easily dealing with a staggering billion pixels of data. These billion pixels are not just random noise; they encode an entire medical library’s worth of information, replete with subtle clues and critical indicators. Furthermore, real-world radiology is fundamentally defined by its edge cases – those rare but often critically important pathologies that appear infrequently but demand immediate and precise identification. We learned a harsh but invaluable truth early on: AI models that perform admirably in the sanitized, controlled environments of research often falter, or even catastrophically fail, when confronted with the boundless complexity and variability of the real clinical world.

To illustrate this challenge, think about the evolution of self-driving cars. A decade ago, the progress seemed breathtakingly rapid, with impressive demonstrations and bold predictions. Yet, as these vehicles moved from controlled test tracks to public roads, the real world relentlessly introduced an unending stream of unforeseen failure modes: unpredictable human behavior, diverse weather conditions, construction zones, unusual road markings, and myriad other anomalies. After more than a decade of immense capital investment, countless hours of testing, and sophisticated technological advancements, only a handful of companies have managed to achieve anything resembling true reliability in autonomous driving. This arduous journey underscores the profound difficulty of translating a technologically impressive prototype into a safely deployable, robust real-world solution.

Components Required to Build Reliable Models: The Integrated System Advantage

As we observed the trajectory of self-driving technology, key patterns emerged that offered profound insights for our own endeavor in healthcare AI. The companies that achieved the most significant and sustainable progress were those that exerted comprehensive control over the entire system and achieved massive scale early in their development. They owned the vehicles, meticulously designed the sensor stack, developed robust data collection pipelines, created sophisticated simulation environments for testing, and managed the end-to-end deployment infrastructure. This level of deep integration, operating at an unparalleled scale, allowed them to continuously collect rare and challenging edge cases, rapidly retrain their models with this new data, rigorously validate improvements, and safely redeploy updated systems. This continuous feedback loop was their secret weapon.

Radiology AI, we realized, is no different. True, enduring success in the real clinical world demands a similar level of systemic control and scale. It necessitates access to massive, historically diverse datasets that capture a broad spectrum of patient demographics, pathologies, imaging modalities, and scanner types, ensuring generalizability and robustness. Equally critical are live data feeds that continuously surface new, rare edge cases and detect distributional shifts in imaging patterns or disease prevalence, allowing for proactive model adaptation. Beyond data, it requires vast clinical resources – expert radiologists for annotation, validation, and feedback – and robust operational infrastructure. This infrastructure is essential for redesigning complex clinical workflows around AI, engineering systems that perform reliably and consistently at scale, conducting rigorous large-scale research studies (often randomized controlled trials) to generate clinical evidence, securing necessary regulatory clearances (like FDA approval), safely refining models post-deployment, and continuously monitoring their performance in real-time.

Furthermore, the recent advancements in frontier language models have unequivocally demonstrated that continuous, high-quality, and extensive human feedback is the "secret sauce" for making AI models genuinely useful and accurate. This principle holds true with even greater import in radiology. In a future where radiology reports are initially drafted by AI, every single draft must be meticulously reviewed, edited, and ultimately signed off by a human radiologist. These human edits and corrections are not merely quality control steps; they represent extraordinarily high-quality signals that can be leveraged to continuously improve the underlying AI models. This creates a powerful, virtuous flywheel effect: better AI models elevate radiologists’ diagnostic accuracy and enhance their capacity to handle more cases. Improved radiologist accuracy, in turn, increases the quality of future training data and correction signals. Increased capacity allows radiologists to take on additional contracts and serve more patients, which then generates even more diverse data and high-quality corrections, propelling the entire system forward. Access to this unique, correction-rich data at a massive scale is incredibly rare in the AI landscape and is virtually unattainable for a standalone AI startup attempting to build it from scratch.

In Healthcare, Growth Follows Evidence: Building Unassailable Trust

In the healthcare sector, trust is not easily given; it is painstakingly earned through demonstrated excellence. This trust rests squarely on a foundation of proven clinical efficacy, unwavering reliability, stringent security protocols, and rigorous adherence to regulatory requirements. For any health system or large radiology group to adopt technology from a nascent startup, particularly when that technology directly impacts patient care workflows and diagnostic decisions, demands an ironclad body of rigorous, real-world evidence. Without this, adoption is simply not feasible.

Crucially, evidence in healthcare is not generated through small, isolated pilot programs. Such limited deployments, while useful for initial testing, do not provide the breadth and depth of data needed to establish widespread credibility. Instead, robust evidence is meticulously built through sustained performance across diverse clinical sites, varied patient populations, multiple imaging modalities, and a comprehensive range of both common and rare edge cases. If an AI system can prove its worth, its efficacy, reliability, security, and scalability within the operational framework of the world’s largest radiology practice – an entity like Radiology Partners – it simultaneously establishes unparalleled credibility across multiple critical dimensions. This holistic validation provides the undeniable proof points that healthcare organizations demand.

Ultimately, in sectors where human lives are at stake and the overarching goal is to construct something truly enduring and transformative, the most effective and often the fastest way to build it is from within the very system one aims to improve. Our decision to sell Cognita a year after its founding was not a concession or a shortened journey; it was a profound acceleration. It provided us with the foundational resources, the unparalleled data access, the clinical expertise, the operational infrastructure, and the regulatory navigation capabilities that were absolutely essential to deliver on our audacious mission: to significantly increase the world’s access to high-quality healthcare through groundbreaking AI. This strategic integration with Radiology Partners, now as Cognita, the AI business unit of Mosaic Clinical Technologies, is not merely a path to commercial success, but the most direct route to realizing our deeply held vision.

My personal drive, cultivated during my undergraduate studies in physics and electrical engineering, has always been singularly focused on leveraging technology to expand healthcare access globally. Convinced that AI held the most promising key, yet acutely aware of its immaturity for real-world clinical application, I pursued a Ph.D. in AI at Stanford University, specializing in foundation models for radiology. My doctoral work culminated in Merlin, a 3D vision-language model for CT interpretation, whose findings were published in "Nature" in 2026 and widely recognized as one of the most significant papers in the field. This academic groundwork laid the intellectual foundation, but it is through the strategic integration with Radiology Partners that the true potential of this research can be unlocked and deployed at the scale necessary to genuinely transform patient care worldwide.