Builders Lab

This publication describes the design, implementation, and evaluation of a privacy-preserving personal health AI agent that avoids sending personal health data to external servers by using a “separated data planes” approach.

The system is designed to run locally and combine multiple knowledge traditions (including Western clinical knowledge and integrative approaches) while keeping data sovereignty with the user.

Note: This is a personal publication and separate from my professional work.

This GPT-based tool walks homeowners through a structured seismic vulnerability assessment based on construction type, build era, foundation, and local risk factors.

The focus is on translating engineering and inspection concepts into language people can understand — helping them recognize risk and decide when professional evaluation is warranted.

Status: Prototype · Available on request

Garden maintenance is highly contextual — dependent on plant type, local climate, season, and recent care. Most tools reduce this to static reminders.

This project explores how AI can adapt guidance based on real-world conditions and user input, translating observation into timely, actionable recommendations.

• Plant-specific care tracking
• Seasonal and situational recommendations
• Reduced cognitive load for day-to-day decisions

Status: Personal prototype · Not publicly released

This project explores when smaller, task-specific language models outperform large general-purpose models in reliability, latency, cost, and explainability.

• Model selection scorecards
• Local vs cloud deployment tradeoffs
• Data quality checks before fine-tuning
• Operational stability over novelty

Status: Internal exploration · Not public

Scholar Assist supports SAT preparation by adapting explanations to where students struggle, rather than pushing more practice problems.

The system emphasizes concept clarity, feedback loops, and guided reasoning instead of volume-based drilling.

Status: Prototype · Not publicly released