Aether
Aether is an AI-assisted medical bill advocate focused on helping patients understand, challenge, and resolve hospital bills with clear next actions.
Inspiration
Unexpected hospital bills are one of the most common sources of financial stress in the United States. Many patients do not know they can request itemized statements, challenge suspicious charges, or apply for hospital financial assistance and charity care programs.
We built Aether to close that gap. The idea was simple: take a stressful, opaque process and turn it into a guided, step-by-step workflow with plain-language explanations, policy-aware recommendations, and practical negotiation materials people can use right away.
Our pilot flow is designed around real-world behavior we observed in patient support contexts, with early emphasis on HCA/TriStar-style billing and assistance policy patterns, while keeping the architecture extensible to other hospital systems.
What it does
Aether provides an end-to-end bill support workflow designed to demystify complex medical coding and financial policies:
AI intake and triage
- Captures core context (hospital network, insurance status, incident summary).
- Establishes a baseline timeline of care to cross-reference against billing dates.
Bill upload and parsing
- Accepts PDF/image uploads with client-side preprocessing.
- Extracts structured bill information in a guided session flow, utilizing robust OCR to accurately pull out CPT codes, service dates, and facility fees.
Itemized bill analysis and error detection
- Normalizes line items and translates complex medical jargon into plain English.
- Compares charges to benchmark/fair-price data.
- Flags potential overcharges, upcoding, suspicious facility fees, and duplicate-like entries.
Income bracket and assistance qualification
- Uses household context and specific hospital policy data to estimate likely assistance eligibility (e.g., Charity Care provisions under 501(r)).
- Calculates potential sliding-scale discounts based on Federal Poverty Level (FPL) guidelines.
Negotiation strategy generation
- Produces a concrete, step-by-step action plan tailored to the specific hospital's billing department.
- Generates customized call script content and supporting guidance (what to say, who to ask for, and how to escalate).
Resolution and tracking
- Records outcomes and computes reduced amount/savings when available.
- Surfaces a session timeline through guided state transitions to keep users motivated and aware of their progress.
Document generation
- Provides downloadable PDFs for strategy and call script artifacts so users have offline reference materials during actual phone calls.
How we built it
The project is implemented as a full-stack TypeScript application, prioritizing data integrity and predictable AI outputs:
Frontend
- Next.js 16 App Router + React 19 for a highly responsive UI.
- Modular chat dashboard with staged workflow modules (upload, analysis, eligibility, strategy, docs, resolution).
- Session info/checklist panel and strategy panel with progressive reveal behavior to avoid overwhelming the user with information all at once.
Backend/API
- Route handlers under
app/apifor sessions, chat, uploads, analysis, assistance, strategy, hospitals, and resolutions. - Orchestrator-driven state machine that advances session steps deterministically. This guarantees the workflow cannot skip critical steps—like income verification—before generating a strategy.
- Workflow guidance layer that produces assistant responses strictly aligned to the current step and verified facts.
- Route handlers under
Data + modeling
- MongoDB with Mongoose models for sessions, messages, uploads, parsed bills, analyses, policies, plans, and resolutions.
- Seeded demo data for repeatable local development.
- Strict data schemas to ensure sensitive information is handled cleanly and predictably.
AI + processing
- AI SDK integrations for LLM-assisted guidance, heavily constrained by prompt engineering to prevent medical or legal hallucinations.
- Textract parsing for robust document extraction, even on low-quality phone photos of paper bills.
- Deterministic pricing/benchmark analysis pipeline to keep core outputs fully auditable (if a charge is flagged, the system knows exactly why).
Challenges we ran into
- State synchronization across panels and modules.
- We had to prevent duplicate module rendering, stale facts, and regressions where completed checklist items appeared to revert. This required building a robust single source of truth in the session state.
- Designing progressive disclosure without confusion.
- Strategy and eligibility outputs needed to appear as actionable modules, not just text summaries. Balancing a conversational AI interface with rigid, reliable UI elements was a major design hurdle.
- Making failures understandable.
- Infrastructure issues (for example, database connectivity/allowlist problems) required clear, actionable error mapping instead of opaque 500s. We also had to handle OCR failures gracefully by prompting users for manual clarification.
Accomplishments that we're proud of
- A complete guided flow from intake to strategy to resolution, firmly grounded in deterministic session state rather than relying solely on LLM context windows.
- Deterministic line-item analysis with benchmark-based flagging and overcharge estimates.
- Policy-aware assistance qualification integrated directly into next-step guidance.
- Practical deliverables users can immediately use (call script, strategy draft, downloadable PDFs).
- Significant UX polish in module sequencing, checklist progression, and overlay panel behavior.
- Robust API error handling that surfaces meaningful status codes and failure reasons.
What we learned
- In high-stress domains like medical debt, trustworthy UX matters as much as model quality. Users need to feel like they are in control of the process.
- Deterministic state machines and typed contracts reduce ambiguity in AI-heavy products and prevent the assistant from going off the rails.
- Progressive module rendering must be tied to real completion signals, not optimistic assumptions.
- Clean error semantics (especially infra errors) dramatically improve developer and user debugging loops.
- Building advocacy tools requires combining technical correctness with deeply empathetic communication.
What's next for Aether
- Expand policy coverage beyond the initial pilot health-system assumptions to cover the largest hospital networks in the US.
- Add richer status/timeline auditing with exportable interaction history.
- Improve document generation into full dispute-letter and application-ready templates for specific state assistance programs.
- Add an "Advocate Mode" tailored for social workers and patient advocates managing multiple cases and client profiles.
- Introduce stronger explainability on why each line item was flagged and exactly how target negotiation amounts were chosen.
- Improve mobile responsiveness and accessibility for lower-friction real-world use, particularly for users whose only internet access is a smartphone.
- Add production observability, monitoring, and outcome analytics for continuous product improvement.
Log in or sign up for Devpost to join the conversation.