OpenHealth: Creating transparency in healthcare

Inspiration

Like any project, we wanted to build something that solved a massive challenge we faced in our own lives. Medical billing in the United States is purposefully confusing, and we realized this firsthand when one of our teammates got slapped with a $202.17 bill for a routine EKG at a local hospital. A cost estimator and a support agent had predicted a $30.77 liability, but the hospital sneaked in a split charge for "facility" and "professional" fees without any warning.

With nearly half of insured adults receiving bills for care they thought was covered, and up to 80% of medical bills containing errors, we realized people aren't just overpaying, but lack the tools even to know when they are being overcharged. We built OpenHealth to bridge this gap between opaque pricing and the patients who pay the price.

What It Does

OpenHealth is a medical bill transparency platform that empowers users to fight back against confusing hospital charges by leveraging autonomous, reasoning AI.

• Multi-Modal AI Bill Auditing: Users can upload a medical bill PDF or image to contribute to community data. Our visual-AI parses the document, auto-extracts bill details, flags potential overcharges, and explains complex codes in simple terms.
• Agentic Data Analysis & Chat: Users can chat with our specialized AI agent to decode their charges and find negotiation points. More than just a chatbot, this agent utilizes real-time tool-calling to query our live database and web search to answer complex questions like, "What is the median cost for an MRI in my city across different insurance plans?"
• CPT Price Search & Cost Insights: Users can search for procedures by plain language or 5-digit CPT codes, utilizing advanced filtering (provider, plan, location, date) to compare "Billed" vs. "Allowed" amounts. Results feature interactive charts, price distributions, and trending procedure analytics.
• Crowd-sourced Data: Users can upload procedures to contribute to openly accessible healthcare costs.
• Personalized Estimates: Users can save an insurance profile to see personalized results and estimate deductibles and out-of-pocket costs based on their specific plan.

How We Built It

We built a robust, real-time stack designed to showcase true agentic behavior and high-fidelity data handling:

• The Agentic Infrastructure (agent.ts): We utilized the ReAct (Reason → Act → Observe) pattern to build a Sandboxed Analyst Agent. The agent uses a Schema Discovery Protocol to understand our Convex tables in real-time, intelligently deciding when to retrieve information. It executes Bounded Queries via strict function-calling to fetch data, ensuring 100% factual accuracy and preventing hallucinations.
• Multi-Modal Agent Pipeline: We implemented a vision-to-JSON pipeline utilizing Nemotron-3-nano-omni-30b-a3b-reasoning and UploadThing for secure temporary document storage. This model combines visual analysis with logical decision-making to differentiate between confusing billing terms (e.g., "Contractual Adjustment" vs "Patient Responsibility") and extract a rigid JSON structure.
• High-Performance Architecture & Admin Controls: To handle our massive crowdsourced medical dataset, we implemented Strategic Denormalization in our Convex schema. We also built secure admin tools for database hydration and synthetic data generation to support our agent evaluation pipelines.

Challenges We Ran Into

Mapping disparate medical concepts (like CPT codes, units, and allowed vs. billed amounts) into a unified, queryable schema was a massive structural headache. None of us is a medical billing expert, so understanding the US healthcare hierarchy was incredibly difficult.

Additionally, ensuring agentic reliability with sensitive financial data is tough; general LLMs are prone to hallucinating specific pricing. We overcame this by leaning into Nemotron's advanced reasoning capabilities and building a rigid validation layer. We forced the agent into multi-step workflows where it must mathematically prove its answers via structured database tool-calling. Finally, ensuring that these complex analytical queries remained fast and responsive on the backend was a significant optimization hurdle.

Accomplishments That We Are Proud Of

We are incredibly proud of our agent.ts implementation. We're proud of it demonstrating true agentic behavior by carrying out actions such as planning, tool-use, and autonomous problem-solving to act as a sophisticated bridge between raw data and natural language. We successfully generated and managed a professional-grade medical dataset that accurately reflects real-world US healthcare pricing standards. Going from being frustrated about insurance overcharges to having a fully functional, multi-modal price-comparison agent within 24 hours is something our whole team is stoked about.

What We Learned

• "Price Transparency" is a legal requirement that hospitals try very hard to hide, and CPT code hierarchies are as complex as any programming language.
• We gained deep experience in multi-agent orchestration and "sandboxing" LLMs, learning how to give them enough power to be autonomously useful while maintaining strict data integrity.
• We learned how to design tool-calling applications where the AI intelligently interacts with external APIs rather than just generating text.
• We discovered how incredibly powerful Convex's real-time database triggers and analytical functions are when paired with an agentic workflow.

What Is Next

• Multi-Agent Negotiation System: A team of specialized agents (Analysis Agent → Strategy Agent → Writer Agent) that collaborate to generate custom appeal letters and negotiation scripts tailored to the specific overcharges found in your bill.
• Map Visualization: A geo-spatial tool to visualize "Medical Deserts" and identify regional price gouging.
• Web Scraping Engine: Automated ingestion and data hydration of the large Machine Readable Files (MRFs) that hospitals are now required to publish.

Tech Stack