# Project export: ReferralGuard This document was generated by HackStack to give an AI agent context about a hackathon project. Sections are labeled with their provenance; content marked as truncated was cut to keep this document small. ## Project metadata - Hackathon: UC Berkeley AI Hackathon 2026 - Tagline: ReferralGuard: multi-agent AI that catches prior-auth denials before you submit - reads typed, faxed or spoken referrals, flags each risk with a reason, auto-submits the clean ones. - Devpost: https://devpost.com/software/referralguard - GitHub: not linked - Video: https://www.youtube.com/embed/T_O8nwm4qO8?enablejsapi=1&hl=en_US&rel=0&start=&version=3&wmode=transparent - Team: contributor stats unavailable ## Devpost submission (written by the team) ### Overview 💡 ### Inspiration A friend doing residency mentioned something that stuck with us. The hardest part of her week wasn't medicine. It was insurance paperwork. The numbers back it up: physicians spend ~13 hours a week on prior authorizations, and roughly half of specialty referrals never get completed, not because the care was wrong, but because a form was missing an ICD-10 code, a member ID was unreadable off a fax, or a payer's step-therapy rule wasn't documented. The maddening part is that almost every one of those denials is predictable. The information needed to know "this will bounce" is sitting right there in the referral. Nobody just checks it before hitting submit. So we built the thing that does. 🛡️ ### What it does ReferralGuard is a multi-agent pre-flight check for specialty referrals and prior-auth requests. Drop in a referral (typed, faxed, or spoken down a phone line) and in seconds you get: a verdict in plain English: READY TO SUBMIT, NEEDS INFO, or HIGH DENIAL RISK a reason for every flag (e.g. "United requires a documented methotrexate trial, not in the note") a live, replayable audit trail of what each agent checked and why a grounded case chatbot that answers only from that referral, so it won't invent clinical facts an auto-filled prior-auth form you can download as a PDF the ROI: time and denial cost saved, per referral, per day, per month If a request is clean, a human approves and a real headless browser submits it to the payer portal and returns a confirmation number. 🧠 ### How we built it It's a real agent pipeline, not a prompt wrapper. Each agent emits one timestamped step: Intake → Extraction → Completeness → Validation → Denial-Risk → Decision → Approval → Submission Claude: field extraction, denial-risk reasoning, the grounded case chat Redis: inter-agent session state and the audit trail Sentry: captures the failure paths (e.g. an unreadable member ID) with breadcrumbs Deepgram (nova-3): transcribes spoken phone referrals Arize Phoenix: OTEL traces of every agent decision Browserbase: drives the live payer-portal submission over CDP Orkes Conductor: the human-approval gate prior auth legally requires Fetch.ai / ASI:One: ReferralGuard is also a discoverable uAgent. Describe a referral in plain English in ASI:One and get the verdict back with no UI, plus an agent-to-agent handoff. The Token Company: the chat compresses its own context before every model call The compression is real and measured with the model's own tokenizer: Reduction = 1 - (T_after / T_before) ≈ 0.5 A Claude guard then verifies every decision-critical field survived compression and restores it if not, so we cut tokens with no quality loss. The ROI panel is honest arithmetic, live on a slider: Monthly Savings = D × (r_time × c_staff + p_deny × c_rework) × W where: D = referrals per day r_time = hours saved per referral c_staff = staff cost per hour p_deny ≈ 0.15 (first-pass denial rate) c_rework ≈ $43 (rework cost per denial) W = working days per month 🧗 ### Challenges we ran into Bleeding-edge Python. We were on Python 3.14, and half the SDKs had no wheels yet. greenlet wouldn't compile, which blocked Playwright; a newer pin fixed it. The Browserbase bug that looked like success. Submissions silently fell back to "retry" with no confirmation. The cause: Playwright's sync API can't run inside FastAPI's asyncio loop. Moving the pipeline into a threadpool made the real browser submission complete. Fetch.ai on 3.14. The uAgent crashed at startup because uagents relied on the implicit event loop Python 3.12+ removed. Creating a loop before building the agent fixed it. Keeping it honest. Everything degrades to a clearly labeled mock with no keys, but each engine is a real SDK call when live, and we made the "live vs mock" line explicit so judges can trust it. 📚 ### What we learned "Looks like it works" and "actually works" are different projects. The most valuable hours went to chasing two silent failures end to end. The best agent UX for a non-technical user (a doctor) is plain language plus a reason for every decision. The audit trail is the product. Context compression isn't just a cost trick. With a verifier guard it keeps agents fast and faithful. 🚀 ### What's next Real payer-policy feeds instead of our 3-rule knowledge base, EHR and fax connectors, and a fine-tuned small model as the dedicated context compressor. ## README (from the GitHub repository) No README available. ## Detected evidence (automated analysis) No repository was indexed for this project. Claimed technologies below could not be checked against code. - CSS (language) — claimed on Devpost, not found in the code - FastAPI (technology) — claimed on Devpost, not found in the code - HTML (language) — claimed on Devpost, not found in the code - JavaScript (language) — claimed on Devpost, not found in the code - Python (language) — claimed on Devpost, not found in the code - Redis (technology) — claimed on Devpost, not found in the code ## Codebase structure No repository index available. ## Key source files No repository index available; no source files included.