Publications

This paper presents the DermUnbound conceptual framework for physician-deployed, privacy-first clinical AI in dermatology. We argue that the current paradigm of cloud-dependent AI services introduces unacceptable risks to patient data sovereignty and clinical autonomy, and propose a model in which physicians directly control the deployment, configuration, and governance of AI tools within their own clinical environments. The framework addresses the technical, ethical, and regulatory dimensions of transitioning from vendor-controlled to physician-controlled AI infrastructure.

We systematically evaluate major cloud-based AI platforms for their applicability to clinical dermatology workflows, examining diagnostic accuracy, integration complexity, and data handling practices. Our analysis reveals significant limitations in cloud-dependent approaches, including inconsistent performance on specialized dermatological tasks, opaque data retention policies, and vendor lock-in patterns that restrict clinical flexibility. These findings support the development of locally-deployed, open-source alternatives that preserve physician autonomy and patient privacy.

This paper describes a technical framework for deploying open-source AI models on local hardware using Docker containers, enabling air-gapped operation and full data sovereignty. We detail the architecture of a containerized clinical AI toolkit that runs dermoscopy classifiers, documentation generators, and pathology processors entirely on physician-owned infrastructure. Benchmarking demonstrates that modern consumer-grade hardware can achieve clinically acceptable inference times while maintaining complete isolation from external networks.

We introduce the concept of vibe coding as a paradigm for clinician-driven software development, demonstrating how practicing physicians can leverage AI-assisted coding tools to build, customize, and maintain clinical applications without formal programming training. Through case studies drawn from the DermUnbound project portfolio, we illustrate how a single clinician developed a suite of production-grade tools spanning surgical decision support, clinical documentation, and pathology report processing. This approach democratizes clinical software development and enables rapid iteration on tools that directly address practice-specific needs.