BillRadar

An AI agent that reads the whole pile of hospital bills and finds the money a family shouldn't have to pay.

Comment

Measured on Token Factory: Nemotron 3 Ultra flagged 6 of 8 wrong findings vs 0 of 8 for Super; Nemotron 3.5 Lightning translates at 53/56 vs 54/56 for Super, about 5.5× cheaper. And it checks the whole pack of papers together: the hospital's bills, the doctors' bills and the insurer's statements, not one bill at a time.

I gave birth twice at the same hospital. In 2024 I paid $7,683. In 2026 I paid $847. The difference, more than $6,800, came down to one question. The second time, my small Telegram bot found a line about financial assistance in the bill. I didn't believe it, so it showed me the law. I was approved for an 83% discount. The same line was in the 2024 bill, and I never saw it. Dollar For estimates that every year US hospitals bill patients at least $14 billion that should have been waived as charity care (Dollar For, "Bridging the Chasm", 2024: https://dollarfor.org/wp-content/uploads/2024/04/Dollar_For.Bridging_the_Chasm.pdf).

After a hospital stay, a family gets a pile of paper: bills from the hospital and the doctors, the lab, letters from the insurer. Errors hide between the pages. The same anesthesia is billed twice. The newborn's care is billed under the mother's insurance ID. And the financial assistance application has a deadline.

BillRadar works on any hospital bill: after surgery, an ER visit, a checkup or a birth. The example here is one case, a birth.

What the agent does

You photograph the table with all the papers (plus any PDFs), and BillRadar works through the case on its own:

  1. It finds every sheet in the photos and straightens it.
  2. On the device, it covers names, member and account numbers, addresses with the apartment and ZIP+4, claim and statement numbers, mailing lines and barcodes before anything goes to a model. Each covered sheet is checked again on the device before it leaves, and a sheet where that check still finds one of these is not sent.
  3. It reads every line exactly as printed and links pages from different letters into one case.
  4. It checks the case against the hospital's own financial assistance policy and federal rules, quoting each rule with its page.
  5. If the hospital is not in BillRadar's base, it finds that hospital's own financial assistance policy with Tavily. It searches only the hospital's own site: the site printed on the bill, or, if the bill has none, a site found by name and city that shows the same name and city or ZIP. When in doubt it asks "Is this your hospital?". Nemotron 3 Super copies the income tiers and discounts, and code keeps a tier only if its quote is in the policy text and the quote itself prints the tier's percentages. Anything unconfirmed is shown as a link: "we found the policy, check it yourself". When the page has no date, BillRadar says it may be out of date.
  6. Code recomputes every amount and date and holds back anything it can't confirm. Then Nemotron Ultra reads the whole case again and flags what doesn't add up. A flagged finding isn't hidden; it moves to "worth checking" without a dollar amount.
  7. It drafts the letters to sign and fills in the hospital's real financial assistance form. Only the family sends anything.

What the family sees

The case screen shows each finding with its sheet and line, the amount, and a quote from the law or the hospital's policy. Next to it are the letters, the filled form and the deadline. Below the case is a receipt: which Nemotron model ran each step and what it cost. For this case: 15 model calls for reading and checking, about 15 cents ($0.146; translating the letters is extra).

For other AI agents

Other AI agents, such as a family assistant, can use BillRadar the way a family does, over two standards: MCP and A2A (/mcp, /a2a and /.well-known/agent-card.json). The agent sends a pack of bills and statements and gets back every finding with the quote and page it rests on, the protections that apply, the letters, the call cheat sheet and the deadlines. The hosted demo is a read-only copy, so the agent entry works when you run BillRadar from the code; the "For AI agents" page shows how to connect.

Why Nemotron, and why different sizes

The models read and explain. Plain code decides whether there is money. Each step gets the model size it needs, chosen by measuring on fictional bill packs:

  • Nemotron 3.5 Super VL sees the photos: it finds the sheets and reads them (up to 8 images per request).
  • Nemotron 3.5 Lightning translates the letters into Spanish and Russian: 53 of 56 parts correct versus 54 of 56 on Super, at 5.5 times lower cost.
  • Nemotron 3 Super answers questions about the case, only with a quote that can be found on a sheet: code first picks candidate quotes from the sheets and the model answers from them, 8 of 10 questions under our strict quote check, a small set (Lightning got 1 of 10).
  • Nemotron 3 Ultra does the second read of the case: it doubted 6 of 8 wrong findings (Super doubted none), and also 4 of 9 right ones. A doubted finding stays on the page under "worth checking", without a sum.

Every call is logged with Token Factory's response (which model answered, the request id, the tokens), so a judge can check what ran.

Where Tavily comes in

Every hospital has its own financial assistance policy, and without it you cannot tell a family whether a discount applies. Tavily finds the document on the hospital's own site only (search with include_domains and the page text in the answer; extract as the fallback), Nemotron reads it, code checks it. It works for hospitals in any state that post their policy. Tested on 5 real hospitals in 4 states: the policy was found for all five. One was confirmed in full with a revision date: Memorial Regional, 5 of 5 tiers, exactly as in our hand-checked base. For three, only some tiers could be confirmed or the page has no date, and BillRadar labels them so. One page has no tier table, so BillRadar shows the link. In the same run BillRadar set aside Form 990 tax returns and community needs reports, Korean, Russian and Spanish versions, a 2018 policy, and Wikipedia and directory pages instead of the hospital's site. 20 Tavily credits for the five.

What it can't do yet

  • Token Factory takes 8 images per request, so big piles are read in groups.
  • There is no speech model there, so notes after a call with the hospital are typed for now.
  • Policy search works only when the hospital posts its policy on its own site, and best for a dated tier table. Sliding scales (like Cedars-Sinai's for 401–600%) are not read yet; BillRadar sends the family to the policy itself.
  • The family types the hospital's city and site from the bill on the case screen; the reader does not copy them yet.
  • State laws: Florida's are checked by hand. For other states BillRadar looks for the law only on the state's official site and shows its exact words with the revision date, marked "may apply to you; check". In the Oct 29 run this found two California sections (§ 127435); for Texas the law page came back without text, so BillRadar gives only the link; for New York the search did not find the law.
  • Every bill in the demo is fictional. In this project, no real bills have been sent to any model.
  • The example family and Seaside Regional Hospital are fictional. The rules and the financial assistance form are the real public ones of Memorial Regional Hospital in Florida, applied to the fictional case.

What's next

  • Read the bill tables on the device itself, so the model only explains what the device has already read.
  • Find financial assistance policies for more hospitals and more states.
  • Voice: speak your notes after a call with the hospital, and get live hints during the call, once Token Factory has a speech model.
  • Test on real families' bills, with their consent.

Try it

Demo: https://billradar-nemotron.retailbox-automation.com/ · code: https://github.com/NadyKor15/billradar-nemotron · the example runs without a key: it replays a recorded live Nemotron run on fictional data.

Four steps, about a minute:

  1. Open the demo and click "See an example". A finished family case opens: 14 sheets, 3 files, and the amount that may come off the bills.
  2. Look at the findings. Each one shows the sheet and lines with the error, the law or hospital policy section behind it when there is one, and which Nemotron model found and read the sheet. "What these lines mean" explains the lines in plain words.
  3. Then open "Letter: request itemized bill" and "Call cheat sheet": a ready letter to the hospital and notes for the phone call.
  4. Open "Find your hospital's policy" (/policy-finder): a recorded live run of Oct 29, 2026 for 5 real hospitals from 4 states. Policies change, so each card carries the run date and "check the hospital's site".

Built With

  • fastapi
  • nebius-token-factory
  • nemotron-3-super
  • nemotron-3-ultra
  • nemotron-3.5-lightning
  • nvidia-nemotron-3.5-super-vl
  • opencv
  • pydantic
  • pypdf
  • python
  • tavily
Share this project:

Updates

Submission history