The future of healthcare is undeniably linked to how we manage and interpret vast amounts of data. Imagine transforming complex medical documents into instant, actionable insights. This wasn’t just a dream at the recent Caltech Longevity Hackathon; our team turned it into a tangible reality, building an AI medical analyst in a single weekend. The goal? To revolutionize how we approach longevity care, making it smarter, faster, and more accessible.
The Urgent Need for AI in Longevity Care
Why focus on longevity? This specialized field thrives on longitudinal data. Patients accumulate extensive records over years, from detailed lab panels and imaging results to critical clinical notes. Navigating this sea of information manually is not only time-consuming but also prone to human error, hindering proactive health management.
Clinicians and individuals alike desperately need fast, explainable triage systems. They want to quickly pinpoint what’s abnormal, identify crucial changes over time, and determine what educational resources to explore next. Our mission was to create a generalizable pipeline that could be built rapidly, even “in a weekend,” proving the immediate impact of AI on real-world medical challenges.
Building Our AI Medical Analyst: Architecture and Workflow
At the hackathon, we didn’t just conceptualize; we shipped a fully functional prototype. Our AI medical analyst features a user-friendly Next.js/React UI with a straightforward uploader and a clean results table. This design prioritizes ease of use, ensuring that anyone can quickly upload a document for analysis.
The system supports mixed inputs through client-side text extraction for PDFs and images, accelerating the initial processing phase. The core intelligence comes from a structured LLaMA prompt, designed to return a comprehensive summary, key medical keywords, relevant categories, flags for abnormal values, a suggested filename, and even PubMed article titles for further research. All raw files are securely stored in Supabase Storage, while structured metadata is managed in a Postgres table.
A Dual-Path Processing Architecture
Our architecture incorporates two distinct processing paths to maximize flexibility and scalability. The client-led path offers immediate feedback, perfect for quick demos and smaller files, processing data directly in the browser.
For larger documents or batch workflows, we developed a server-led path utilizing a Supabase Edge Function. This approach ensures scalable, secure, and robust background processing capabilities. The overall workflow is simple: Upload → Extract text → LLM analysis → Persist → Render, providing a seamless user experience.
Key moving parts in our stack included Frontend: Next.js + Tailwind; OCR/Parsing: pdf-parse, tesseract.js; AI: LLaMA chat completions API with its rigid, parse-friendly prompt; Backend: Supabase (Storage for blobs, Postgres for metadata); and Serverless: Supabase Edge Function for server-side PDF processing.
The Prompt That Powers Precision
The intelligence of any large language model (LLM) hinges on its prompt structure. To ensure reliable and predictable parsing, we engineered a highly structured LLaMA prompt. This rigid schema minimizes brittle, free-form responses, making the output consistently actionable.
The prompt we devised forces the LLM to adhere to a specific format, instructing it to provide a clear, plain-English summary, extract key medical terms and their values, classify content into predefined categories, suggest a descriptive filename, identify abnormal values with “high,” “low,” or “normal” flags, suggest relevant PubMed article titles, and include any additional medical guidance or observations.
For example, the prompt instructed the LLaMA model with explicit formatting requirements:
Summary: [summary]
Keywords: [key:value pairs]
Categories: [categories]
Filename: [filename]
Flags: [abnormal values]
References: [article titles]
Notes: [additional guidance]
This rigid structure was crucial for the reliable extraction of data points.
Beyond the Hackathon: Key Considerations and Future Horizons
The hackathon prototype provided a strong foundation, but a production-ready AI medical analyst requires addressing several critical areas. UX considerations included designing a drag-and-drop uploader with clear accept types, providing visible progress and error states, and presenting terse, readable summaries with expandable details. We also planned for badges for categories and clear flags for abnormalities.
Reliability strategies centered on the structured prompt for predictable parsing, keeping LLM temperature moderate (0.3–0.7) to reduce variance, and validating parsed JSON fields with safe fallbacks. We also tracked versions and statuses to support re-processing and migrations, ensuring data integrity over time.
Security and compliance are paramount when dealing with sensitive health data. This means treating all uploads as potentially sensitive, never exposing secrets client-side, and considering de-identification or redaction at upload. Encryption at rest (handled by Supabase storage), HTTPS for all calls, Row Level Security (RLS) across documents, and signed URLs for downloads were all key security considerations.
Looking ahead, our vision extends to normalizing lab values with medical ontologies like LOINC and unit conversions, enabling trend analysis and change detection over time. Integrating confidence scoring and a reviewer checklist would enhance clinical safety, while human-in-the-loop editing with audit trails would ensure accuracy and accountability. Ultimately, exporting data to FHIR-compatible bundles would pave the way for seamless integration into existing healthcare systems.
Conclusion
Building an AI medical analyst in a weekend at the Caltech Longevity Hackathon demonstrated the incredible potential of focused innovation. By combining powerful AI with a robust serverless infrastructure, we created a tool that can quickly transform complex medical paperwork into fast, explainable insights. This project is a testament to how rapidly new technologies can be leveraged to address critical needs in healthcare, paving the way for a future where patient longevity is supported by intelligent, accessible, and precise data analysis.
