AeonCura

Inspiration

Ten weeks embedded with specialty-clinic coordinators showed us the revenue cycle's dirty secret: whether a $5–20K infusion claim survives is decided before the visit, across five systems that don't talk — yet the entire healthcare-AI market automates downstream, after the denial. Our design partner had even bought a prior-auth bot and shelved it: wrong end of the pipe.

What it does

AeonCura's Chart Prep agent works a clinic's schedule every morning — querying eligibility, PM, EHR, labs, balances, and credentialing; reconciling conflicting sources; and scoring every patient green/yellow/red with a human clearing each check, drafted outreach on yellows, and dollar-quantified escalations on reds. It's the first module of an EMR-agnostic MSO operating platform that attacks revenue leakage upstream: same connector layer, next modules are referral intake, PA renewal, and drug routing.

How we built it

A supervisor agent fans LLM-powered workers across six (synthetic but behaviorally faithful) clinic systems via a tool-use loop, with every output crossing a typed pydantic contract, a deterministic guardrail gate that can only downgrade severity — the LLM proposes, code disposes — and a hash-chained audit log, streamed live over SSE to a coordinator dashboard. An eval harness tracks our headline safety metric, false-green rate, and fails CI if a single at-risk patient slips through.

Challenges we ran into

Making a probabilistic agent safe for a domain where one false-green means a five-figure claim writeoff — solved by moving safety rules out of the prompt and into a deterministic policy gate downstream of the model. And faithfully simulating how clinic systems actually fail: eligibility feeds that return silence, network truth living in one person's spreadsheet, insurance changes arriving as scanned cards nobody re-keys.

Accomplishments that we're proud of

Zero false-greens across an 18-patient eval - enforced in CI, plus a tamper-evident audit chain you can verify live at /api/audit_verify. And a demo where you can watch the agent hit a silent eligibility response, reroute to a credentialing fallback, catch an expired prior auth colliding with a payer-identity conflict, and refuse to resolve it alone.

What we learned

In regulated domains, the product isn't the agent — it's the evidence: every finding needs a source, a confidence, a timestamp, and a human who cleared it, because that's what makes an operating clinic group say yes. And upstream beats downstream everywhere: ten minutes of agent work before a visit is worth more than any amount of AI litigating the denial after.

What's next for AeonCura

Live pilot with our design-partner clinic group — real Availity, Veradigm, and portal adapters replacing the mocks — then the same connector layer powers referral intake and PA renewal, because every module shares one canonical patient-readiness record. The wedge is chart prep; the company is the operating platform for MSO roll-ups, where every acquired clinic is a pre-sold install.

Built With

• ci-cd
• evals
• fastapi
• github-actions
• jsonl
• lovable
• playwright
• postgresql
• pydantic
• pytest
• python
• railway
• rest-api
• server-sent-events
• tavily
• uvicorn