I build to understand. Not to delegate.

Designed and built a local-first AI agent that keeps personal health data entirely on-device. Went hands-on with privacy-preserving architecture, not to read about it, but to understand the tradeoffs from the inside. The patterns directly inform how I think about data sovereignty in clinical AI deployments.

Built a multimodal AI tool that identifies plants, diagnoses problems, and generates location-aware seasonal maintenance plans from a photo. The clinical application is irrelevant. The exercise in multimodal reasoning, context injection, and non-clinical user experience design was not.

A GPT that generates seismic risk assessments for residential properties based on location, construction type, and local geological data. Built to make technically dense risk information accessible to people without a geology degree. Same challenge as clinical AI: translating expert knowledge for non-expert decisions.

An adaptive AI tutoring tool for SAT preparation. Built to get hands-on with personalized learning outside of healthcare, where the stakes are lower and the iteration cycle is faster. Tested prompting strategies that have since shown up in clinician-facing tools.

Hands-on work with small language model architectures, understanding what they can and can't do on constrained clinical tasks, and what deployment looks like outside cloud-dependent infrastructure. Built because the answer wasn't in a paper.