Inspiration
WhaInspiration
Prior authorization killed a patient in my family's life before we ever understood what prior authorization was. The denial came back in a letter. By the time the appeal was approved, the window for intervention had closed. When I dug into the data building this project, I found out that story is not exceptional — 29% of physicians report PA delays have caused serious adverse events, including death. 94% say it directly contributes to burnout. 82% have watched patients abandon treatment entirely while waiting for approval. The system is not broken at the edges. It is broken at the center, by design, and the people paying the price are patients and the clinicians who chose medicine to help them.
What made it feel solvable was a simple observation: most denials are not about coverage disputes. 80.7% of appealed denials are reversed. That means the evidence was obtainable — the documentation was findable — the claim was approvable. The problem is that clinicians submit against rules that changed last week, with documentation requirements they didn't know had shifted. PolicyPulse exists to close that gap.
What It Does PolicyPulse is a fully autonomous prior-auth intelligence agent. It continuously monitors trusted payer policy sources, detects when something material changes, and waits silently in the background until a clinician action arrives via CDS Hooks. When a doctor signs an order for a procedure whose payer rules just changed, PolicyPulse surfaces a CRD guidance card with a Senso-grounded DTR checklist — the exact documentation required before submission. When the doctor's action has nothing to do with the detected change, the agent stays completely quiet. Every decision, including every suppressed alert, is written to ClickHouse for a full, auditable event trail.
How We Built It We built the agent core first, with no external APIs — just local policy fixtures, a diff engine, and a materiality classifier routing decisions to four outcomes: crd_guidance_ready, suppress, human_review, and log_only. Once the logic was clean and the routing was reliable, we layered in the integrations in dependency order: ClickHouse for event memory and auditability, then x402 for the payment gate, then Senso for grounded checklist retrieval and cited.md publishing, then Nimble for trusted live policy ingestion with a fixture fallback to protect the demo.
The backbone is PolicyPulseService — a single service class that owns the autonomous monitor thread, the scan loop, the CDS Hook ingestion pipeline, and the full readiness card generation flow. The background thread starts the moment the Streamlit UI loads, before the first user interaction, running on a 15-second scan interval. All state flows through ClickHouse, with a dual-write local JSONL fallback so the system runs reliably regardless of cloud connectivity. The Streamlit UI was built to feel event-driven, not form-driven — no "Generate Alert" buttons. The agent decides; the UI reports.
Challenges The hardest challenge was suppression, not detection. It is easy to build a system that alerts. It is hard to build one that earns the right to interrupt by demonstrating it knows when not to. We had to think carefully about what "relevant" means across payer, plan type, procedure code, provider action type, and materiality classification — and make sure that irrelevant triggers genuinely went silent rather than generating noise. That routing logic went through more iterations than anything else in the codebase.
The second challenge was demo reliability. Nimble fetches live payer policy pages, which can return unexpected content structure, timeouts, or empty payloads under demo conditions. We built a layered fallback — live CLI first, fixture fallback second — so the full autonomous loop runs end-to-end regardless of network state. Every integration has a fixture fallback. The demo never depends on a live API call to tell its story.
Accomplishments Getting the suppression path right felt like the real win. Showing judges the "irrelevant doctor fixture" path — where the agent detects a material policy change, an EHR order arrives, but the agent stays completely silent because the payer and procedure don't match — captures what makes this system different from every other prior-auth tool that just adds more alerts to an already overwhelmed system.
What We Learned Healthcare's administrative complexity is not incidental — it is load-bearing for a lot of the revenue cycle. Building a system that intervenes intelligently means understanding that silence is as valuable as a well-timed alert. We also learned that the most important infrastructure decision in an agentic healthcare system is the audit trail. In a compliance-sensitive domain, "the agent decided" is not an acceptable answer. Every decision needs to be traceable, logged, and explainable — which is exactly what the ClickHouse event store provides.
What's Next The hackathon scope targets UHC public policy pages and one realistic scenario. The production path is multi-payer — Myndshft for normalized payer-rule requirements across hundreds of national, state, and regional plans, Availity for authorization workflow context, and Accountable HQ for HIPAA compliance controls and full audit trail support at scale. The x402 payment model also points toward a meaningful business model: charge a fraction of a cent per readiness card generated, and PolicyPulse pays for itself the first time it prevents a denial.t it does
Log in or sign up for Devpost to join the conversation.