Aevora

Inspiration

Healthcare becomes especially frightening when the treatment is over and the bill arrives.

We started with a simple question: what happens when a patient has the estimate, insurance approval, prescriptions, reports and final bill, but no system that can connect them?

India's financial burden makes this more than a paperwork problem. WHO's India SDG profile reports that 17.5% of people spent more than 10% of their household expenditure on health care, while about 5.5% of the population, roughly 78 million people, were pushed into poverty by out-of-pocket health spending. A recent 2026 LocalCircles survey of 23,084 respondents across 306 districts found that 68% of private-hospital users reported excessively high treatment prices, 63% reported excessively high test prices, and 40% reported unexpected or unknown charges.

The emotional part is easy to understand. A family can spend years saving for a major procedure, trust the hospital when treatment is urgent, and only later discover that the final bill looks very different from what they expected. At that point, asking "why was I charged this?" is difficult because the evidence is scattered across different documents.

We found that many existing tools focus on explaining a bill, estimating prices, or finding cheaper alternatives. Those are useful, but we wanted to explore a different question:

Can AI reconstruct the story behind a bill?

That became Aevora.

What it does

Aevora is an AI patient advocacy layer that connects the clinical story with the financial story.

Instead of treating a hospital bill as an isolated PDF, Aevora brings together estimates, final bills, insurance approvals, prescriptions, reports, procedure notes and other treatment evidence.

Its core idea is a CareFlow, an evidence graph that connects:

Treatment event → supporting evidence → estimate → insurance approval → final charge

This lets Aevora reason about changes rather than simply summarize documents.

For example, if a final bill contains a charge that was not present in the original estimate, Aevora can trace the charge through the available evidence and ask:

When did this appear?
Is there a treatment event associated with it?
Is there supporting documentation?
Was it included in the original estimate or package?
Does it differ from the approved amount?
What should the patient ask the hospital about it?

Aevora deliberately does not label a charge as fraud simply because it looks unusual. Instead, findings are classified as Documented, Corroborated, Needs Clarification, Evidence Incomplete, or Review Recommended.

That distinction is central to the product. The AI should help a patient investigate, not manufacture certainty.

The AI reasoning layer

The interesting part of Aevora is the relationship reasoning.

A conventional document AI system might extract:

"Additional diagnostics: ₹17,500"

Aevora tries to understand where that number sits in the patient's larger treatment story.

It can connect the financial entry to the timeline, compare it with the estimate and approval, search the available clinical evidence, identify what is supported and what is missing, and then convert that analysis into a practical next step.

The result is not just an explanation. It is an evidence-backed conversation plan.

Through the Advocacy Center, Aevora can generate questions, scripts, action steps and an evidence pack that a patient can use when speaking with billing teams, insurers or hospital administration.

Aevora also includes a conversational voice interface so the patient can ask questions naturally rather than navigating a financial dashboard for every interaction.

How we built it

We built Aevora as a web application with an AI reasoning layer designed around evidence and traceability.

The application processes healthcare documents, extracts relevant entities and amounts, reconstructs a treatment timeline and maps financial entries to the available evidence.

The main technology stack includes:

Next.js and React for the application
TypeScript for application logic
Tailwind CSS for the interface
Featherless AI for the reasoning and conversational AI layer
ElevenLabs for realtime speech input and voice output
Structured document processing for extracting clinical and financial information

We designed the AI workflow around a simple principle:

Evidence first, explanation second, action third.

The system is also intentionally constrained. It does not diagnose medical conditions, prescribe medication, make legal determinations or guarantee insurance coverage. When the evidence is insufficient, the system should say so.

For our demonstration, we created a synthetic case involving Ananya Sharma, a 47-year-old patient undergoing total knee replacement. The synthetic case contains an original estimate of ₹2,15,000 and a final bill of ₹3,07,400, creating a ₹92,400 difference. Aevora then explores the changes through the treatment timeline, including additional procedure, diagnostic and consumable entries.

The demo data is synthetic by design so that we can demonstrate the workflow without exposing anyone's private medical information.

Challenges we ran into

The hardest problem was not getting an LLM to read a medical bill. Modern models can already summarize documents very well.

The harder problem was asking the AI to reason across documents without turning missing information into assumptions.

A single treatment can appear under different names across a prescription, procedure note, insurance approval and billing system. Dates can also differ, and a final bill may contain information that is impossible to interpret correctly without looking at the treatment timeline.

We therefore had to design Aevora around relationships and evidence rather than keyword matching.

Another major challenge was trust.

In healthcare, an overly confident AI can be worse than an unhelpful one. We had to make sure Aevora could distinguish:

"The document says this"

from

"This appears unusual and needs clarification."

That led us to the evidence-status system and the rule that Aevora should never make accusations from incomplete evidence.

The voice interface was another significant engineering challenge. We wanted a real conversational loop rather than a microphone button attached to a normal chatbot.

The intended flow became:

Microphone → realtime speech-to-text → AI reasoning → text-to-speech → playback → listening again

Synchronizing these asynchronous systems while preventing the assistant from hearing its own voice required careful state management, realtime diagnostics and repeated testing.

Accomplishments that we're proud of

We are proud that Aevora is not just another interface for asking an AI to "explain this bill."

The central idea is the CareFlow evidence graph, where a financial entry is meaningful only when viewed alongside the treatment event, supporting documents, estimate and insurance context around it.

We are also proud of the Financial Drift view, which makes the gap between what was expected, what was approved and what was finally billed visible at a glance.

In our synthetic demonstration:

Original estimate: ₹2,15,000
Insurance approved amount: ₹1,80,000
Final bill: ₹3,07,400
Increase over estimate: ₹92,400
Difference from approved amount: ₹1,27,400

Rather than stopping at those numbers, Aevora traces the change into specific areas such as an additional procedure, additional diagnostics and consumables, then asks what evidence supports each change.

That is the part we believe makes the project meaningful.

We are also proud of the Advocacy Center because finding a discrepancy is only half the problem. A patient still needs to know what to do with that information.

What we learned

We learned that in a sensitive domain, better AI does not necessarily mean more confident AI.

The most useful answer is sometimes:

"I cannot verify this from the documents I have. Here is exactly what you should ask for."

We also learned that document intelligence becomes much more powerful when documents are treated as a connected evidence network instead of separate files.

Most importantly, we learned that patient advocacy is not about replacing doctors, hospitals, insurers or lawyers.

It is about giving the patient enough clarity to participate in the conversation.

What's next for Aevora

Today, Aevora demonstrates the concept using a controlled synthetic case. The next step is making the evidence layer much broader and more reliable in real-world settings.

We want to build a privacy-preserving healthcare transaction benchmark that can help Aevora understand not only what changed, but how a charge compares with relevant historical patterns.

We also want to expand:

Insurance claim and pre-authorization analysis
Package and tariff reasoning
More healthcare document formats
Multilingual and voice-first patient advocacy
Stronger evidence citations for every AI finding
Privacy-preserving storage and processing
Patient-generated evidence packs for disputes and clarification
Longitudinal CareFlow across multiple hospital visits

Our long-term vision is simple:

A patient should not need to become a medical billing expert to understand what happened to their money.

Aevora is our attempt to give patients an intelligent layer between a complex healthcare system and the person who has to navigate it.

Research behind the problem

WHO India SDG Profile: 17.5% of people spent more than 10% of household expenditure on health care, and approximately 78 million people were pushed into poverty because of out-of-pocket health spending.
Government of India's National Health Accounts: out-of-pocket expenditure remains an important measure of the financial protection available to households.
LocalCircles 2026 survey reported by India Today: 23,084 respondents across 306 districts; 68% reported excessively high private-hospital treatment prices, 63% reported high test prices, 40% reported unexpected or unknown charges, and only 14% reported none of the listed billing concerns.
A 2026 longitudinal study of breast cancer survivors in India found that the proportion taking loans increased from 38% at baseline to 65% at the endline and 69% during follow-up, illustrating how prolonged treatment can translate into growing financial debt.