markovHealth

Inspiration

Every year, billions of dollars are poured into life-saving medical breakthroughs that never reach the market, not because the science failed, but because the logistics did. Over 80% of clinical trials fail to meet enrollment deadlines because we expect critically ill patients to act as their own medical researchers and travel coordinators. Inspired by Dario Amodei’s "Machines of Loving Grace," we wanted to build a future where AI does not just process data, but actively removes the physical and cognitive barriers to human health.

What it does

markovHealth is a decentralized clinical trial orchestrator that flips the traditional recruitment model.

For Patients: A seamless "drag-and-drop" portal where Claude 3.5 Sonnet instantly decodes complex genomic reports and medical PDFs into a deterministic match.

For Researchers: A real-time logistics dashboard powered by a Markov-inspired routing engine.

Instead of the patient finding the trial, the cure finds the patient. The system dynamically decides the optimal delivery of care: whether that involves dispatching a mobile phlebotomist, mailing an at-home DNA test kit, or routing the patient to a local hybrid clinic.

How we built it

We built a full-stack solution in a single 8-hour sprint: The Brain, a Python/FastAPI backend that uses Claude 3.5 Sonnet to transform unstructured medical text into structured JSON logic. The Logistics Engine: A custom routing heuristic modeled as a Markov Decision Process (MDP). We implemented a multi-objective reward function to balance patient well-being against operational overhead: $$R = -(w_1 \cdot \text{Friction}) - (w_2 \cdot \text{Cost}) + (w_3 \cdot \text{UrgencyBonus})$$ The Interface: A Next.js and Tailwind CSS frontend featuring an interactive Mapbox dashboard that visualizes real-time routing paths using quadratic Bezier curves for a professional "flight-path" aesthetic.

Challenges we ran into

Our primary technical challenge was implementing the Reinforcement Learning component. Training a deep RL model from scratch within an eight-hour window is unfeasible due to the lack of specialized medical logistics datasets and the time required for model convergence.

We addressed this by engineering a heuristic policy engine that mimics the behavior of a trained agent. We designed a multi-objective reward function to evaluate the state space of patient coordinates and clinical urgency. This allowed us to achieve high accuracy and full interpretability without a month of training.

Accomplishments that we're proud of

As a team of two, we successfully built a fully integrated full-stack application by defining a strict API contract from the start. We are proud of how we decoupled the NLP parsing, the logistics engine, and the Mapbox visualization. This allowed us to work in parallel and integrate our systems seamlessly in the final hour without major merge conflicts or schema mismatches.

What we learned

We gained a deep appreciation for the "Logistics Gap" in healthcare. We learned that the bottleneck is not always a lack of a drug, but rather the friction in the supply chain. Technically, we leveled up our skills in structured prompt engineering and geospatial optimization.

What's next for markovHealth

We want to move beyond mock data to real-world HL7/FHIR medical record integration. Our long-term vision is for markovHealth to become the API layer for every decentralized trial in the country, partnering with mobile health fleets to ensure that your zip code never determines your ability to survive.