Numora Health
24.2M rates. Move your cursor.

Hospital prices, finally readable.

Every Houston hospital publishes the rates it negotiates with insurers. The files run to millions of rows and freeze a laptop. Numora turns them into a search: describe a procedure in plain English and see what each hospital charges next to what Medicare pays, with bad data flagged instead of buried.

Try the searchHow it works
24.2Mnegotiated rates, searchable
50+Houston hospitals
17payers, every rate benchmarked to Medicare
73Mraw records normalized into one dataset
Numora searchSample data. Real rates come from each hospital's published file.
Payer shown: a sample commercial PPO. Flagged rows stay visible and never rank as the cheapest option. Episode totals are labeled as the sum of listed rates, because that is what they are.
Backed, used and supported by

The data is public. Usable is a different thing.

The Hospital Price Transparency rule has been in force since 2021. Hospitals complied in the narrowest way it allows.

What a patient needs

Procedure
MRI brain without contrast
Code
70551
Facility
Academic medical center, TMC
Your plan
Commercial PPO
Negotiated rate
$2,480
Medicare pays
$236
Cheaper nearby
$610, Sugar Land
Data check
Verified

What a hospital publishes

// one row of one hospital's machine-readable file
{"description": "MRI BRAIN W/O CONTRAST",
 "code_information": [{"code": "70551", "type": "CPT"}],
 "standard_charges": [{
   "setting": "outpatient",
   "gross_charge": 5312.00,
   "payers_information": [
     {"payer_name": "PAYER_14", "plan_name": "PPO",
      "standard_charge_dollar": 2480.00,
      "methodology": "fee schedule"},
     {"payer_name": "PAYER_02", "plan_name": "HMO",
      "standard_charge_percentage": 68,
      "methodology": "percent of total billed charges"},
     {"payer_name": "PAYER_09", "plan_name": "EPO",
      "standard_charge_dollar": 14.00,
      "methodology": "other"},
     … 14 more payers
   ]}]}
… 24,199,999 more rows across 50+ files
Left: one row of a hospital's published file. Multi-gigabyte, its own format, codes without descriptions.Right: the same rate in Numora, benchmarked and checked.

The maximum penalty for ignoring the rule is about $2 million a year, and it is rarely enforced. The confusion is worth far more than that to the people it protects.

A family comparing an MRI. A clinic deciding where to refer a patient. A researcher studying the market. None of them can open a file that freezes a laptop. Numora exists to close that gap.

Built so a wrong answer can't look right.

Four decisions that make the numbers trustworthy, and the pipeline behind them.

Verified codes only

The plain-English search is a retrieval-constrained resolver. It can only choose from 18,154 real CPT and HCPCS codes, and if nothing matches it says so. Symptoms never resolve to a procedure.

"mri of my head"→70551 · verified

Reliability flags

Deterministic rules catch placeholders, chargemaster dumps, percent-of-charge misreporting, uniform-rate contamination and commercial rates below 0.7× Medicare. Flagged rows stay visible and never rank first.

MRI brain · $14.00→Likely error · below 0.7× Medicare

Episode estimates

Combine the codes in a procedure into one estimate per hospital, labeled as the sum of listed rates. Hospitals missing a code are marked incomplete and sorted below the complete ones.

45378 + 00812 + 88305=$2,950sum of listed rates

A kill switch on everything new

Every layer added to the product ships with an off switch. Turn the AI off and the product keeps working on codes alone, instantly.

AI search on/off: zero model calls
Step 1Ingest73M

raw records pulled from every hospital's published file.

Step 2Normalize24.2M

negotiated rates, each keyed to code, payer, plan and facility.

Step 3Benchmark17

payers, every rate set beside what Medicare pays.

Step 4Flag7

reliability rules run on every result, in the browser.

Step 5Search18,154

verified codes the resolver can choose from. Never one it invented.

Runs on Palantir Foundry with a React frontend. Supported by the Palantir Valley Forge Grant and research data from Turquoise Health.

Why I built it

Keon Etminan, founder

When my doctor ordered an MRI, I did what the law says a patient can do and went to the hospital's website to find the price. The page moved from soft blues and smiling nurses to a directory of enormous, cryptically named files. The first one I opened froze my laptop.

That contrast stayed with me. The rule required hospitals to publish their prices and did nothing to reward making them readable, so they didn't. I later learned the name for this: mechanism design, the study of how rules shape behavior. Compliance here was a format choice.

I had been handling MRI and CT files since I was fourteen, when I started a small company building imaging exhibits for personal injury lawyers, so the files themselves didn't scare me. I started writing parsers. About six months later, that was Numora.

What I personally built

The parts of Numora that exist because I made them.

  • The parsers

    Wrote the code that reads every Houston hospital's file, and learned medical billing well enough to tell when two differently formatted rows are the same service.

  • The data model

    Designed the 24.2-million-row dataset of Texas hospital charges, every rate keyed to code, payer, plan and facility, and worked through the indexing limits at that scale directly with Palantir.

  • The reliability engine

    Wrote the flagging rules as deterministic checks rather than a model, then moved them client-side so flags compute in the browser on the rows a search returns.

  • The plain-English search

    Scoped it as a retrieval-constrained resolver so the model can only pick from real codes, and shipped it behind a kill switch demonstrated live before launch.

  • The episode estimate

    Rejected two designs that changed the interface too much and settled on one table that sums codes per hospital, labels incomplete rows and lets a flagged component block the total.

  • The first customer

    Brought Numora to Next Level Urgent Care and shaped it around a real workflow: where to send a patient for care a clinic can't provide.

Much of the implementation is done with an AI developer agent inside Foundry. The data model, the rules, the constraints and every product decision are mine, and I review and merge every change.

Where it stands

In use and funded.

In useNext Level Urgent Care

45+ clinics across Texas and 600,000+ patients a year. Uses Numora to decide where to send patients for care a clinic can't provide.

1 of 30Palantir Valley Forge Grant

Selected from applicants nationwide. Numora runs on Palantir Foundry.

DataTurquoise Health

Research data grants supporting the normalization work.

45+Clinics referring with it

Every Next Level location across Texas can route a patient using Numora's comparison.

600K+Patients a year

Seen across the clinic network that routes referrals with Numora.

6 moFrom frozen laptop to product

Built in about six months, from the first parser to a tool a clinic network relies on.

What comes next

The same dataset, pointed at the bill.

The consumer layer for healthcare prices, the way Zillow became the consumer layer for housing.

Keon Etminan

Keon Etminan

Founder, Numora Health. Houston, Texas.

Keon is a senior at The Awty International School in Houston. He founded Total Ready Imaging at 14, which builds medical-legal imaging exhibits for personal injury firms and still runs, and served as a United States Senate Page in spring 2026.

He has done dataset work with the Brown University School of Public Health for an NBER study on hospital IT spending.