TrueCost

Inspiration

Millions of patients each year will not fill their prescriptions due to cost related issues. The people who are most vulnerable are being hurt by our current healthcare system. Surescripts, who run the network most US prescriptions travel across, publish two numbers worth putting side by side: 27% of new prescriptions are never dispensed, and 14.2% take seven days or longer to fill. A lot has been built to detect the first. Almost nothing tells it apart from the second.

What it does

TrueCost helps patients actually fill the prescriptions they've been given. It reads a patient's own cash flow to find the earliest day they can afford their copay (we call it the Ready Day), then works out what kind of help they need. If they can afford it but not until payday, we build a payment plan that fits their real finances. If they can't afford it at all, we check which assistance programs they qualify for and point them there. We want to help everyone along that range, from people who just need a schedule to people who need real support.

On the prescriber's side it does something less obvious: it refuses to send most of the alerts. Of twelve flagged patients, four leave the building and eight do not. The ones who will fill on payday anyway generate no message, because there is nothing for a doctor to do about them. Every refusal is shown on screen, with its reason, next to the ones we sent.

Every patient in our demo is synthetic.

How we built it

Next.js and TypeScript, seven pages and six API routes, deployed at truecost.health. The adjudication is a deterministic engine with no language model anywhere in the decision path, so every figure on screen can be recomputed by hand — 773 assertions check it against arithmetic we worked out ourselves first. Capital One's Nessie supplies the bank side across six endpoint families and two customers: payday cadence, recurring bills and the everyday spending floor are all inferred from transactions rather than typed into a form. The side kick runs that same classifier as a HokieAI agent whose instructions are the classifier, written out as arithmetic it has to show its work on.

Challenges we ran into

Our main challenge was finding a way to meaningfully add functionality to something that already existed. We realized that a large portion of our solution was already implemented by Impiricus, so we had to find a way to innovate and substantially improve upon their implementation. Our answer was to work out the date a patient's own bank account can cover a prescription, and use it to decide when an alert is actually worth sending.

We also found our own demo was wrong, by attacking it ourselves. The headline example had a bill displaced by a payment with payday landing in between, which our own formula makes impossible.

Accomplishments that we're proud of

We are proud for implementing a project that can positively impact millions of Americans. We had never deployed a website before, and this one is live at truecost.health on a real domain rather than only on localhost. We are also proud of what we deleted: a money-moving feature we cut after finding that a sponsor had integrated a company doing exactly that eleven days earlier, and several numbers we could not source.

What we learned

We learned how to effectively communicate as a team and work side-by-side with agentic AI tools. We also learned how to take existing ideas and improve upon them, and that sometimes the most valuable thing a system can do is stay quiet.