Vasant Kearney

Vasant Kearney is the founder and chief executive of Onlay AI, which builds dental revenue-cycle automation spanning insurance verification, claims, clinical attachments, payment posting, and reconciliation. His work applies a medical physicist’s understanding of imaging and clinical uncertainty to one of healthcare’s most stubborn operational problems: reliably coordinating insurers, providers, and patients.

From medical imaging to healthcare operations

Kearney holds a doctorate in biomedical engineering and completed a radiation oncology physics residency at the University of California, San Francisco in 2018. He subsequently served as an assistant professor in UCSF’s radiation oncology department, working on machine learning for medical imaging and treatment planning.

In 2020, he conceived and co-led development of DoseGAN, an attention-guided generative adversarial network that predicts radiation-dose distributions from patient anatomy. Its design addressed a consequential weakness of conventional optimization: reducing pixel-level error can smooth away clinically important features. Kearney later argued publicly that realistic synthetic medical images and familiar evaluation metrics can move in opposite directions.

As co-founder and chief technology officer of Retrace Labs, he extended this research into dentistry. A 2022 study he led used generative dental-image inpainting to reconstruct anatomy beyond the limited frame of dental radiographs, improving estimates of periodontal attachment levels.

At Onlay, Kearney expanded from clinical interpretation into the surrounding administrative workflow, connecting practice-management software, payer portals, insurance transactions, imaging, and financial systems. His goal is to reduce insurance-administration costs while improving the experience for patients and clinical staff.

How he designs reliable healthcare agents

• X12 as an agent execution harness: Kearney uses established healthcare transaction standards to constrain eligibility checks, claims, attachments, status inquiries, and payments. His healthcare-agent framework makes actions inspectable and rejectable without hard-coding every operational variation.

• Standardized does not mean correct: Insurer portals, call centers, and transaction feeds can disagree—or agree on coverage that a subsequent denial disproves. Kearney treats payer information as provisional and revisable.

• Preserve clinical context: Reducing images to isolated findings can discard details needed later for treatment interpretation or claim documentation. Multimodal information should remain available across downstream decisions.

• Bounded organizational memory: Enterprise agents need database-backed separation between organizations, partners, and users, while allowing people to override assumptions inherited from previous tasks.

• Revalidate every model upgrade: Changing models changes application behavior. Kearney emphasizes renewed evaluation, auditable actions, human escalation, and tight control of multistep reasoning costs and error propagation.