Method · Open by design

How a hospital bill becomes a dispute letter in under a minute

Five stages, deterministic where it matters, AI only for reading and narrating — never for citing rates. Below is the entire pipeline, in the same order it runs on your bill.

The five stages of an audit

1
You upload a bill
Drag-drop a PDF or photograph the bill page-by-page. No login. The original file is purged after thirty days; only the PII-scrubbed structured data is kept.
2
We read every line
Claude reads each line item — name, quantity, unit price, total, date — using a strict JSON schema. Low-confidence units are flagged for verification and excluded from any dispute document.
3
We look up the government reference
Medicines are matched against the NPPA ceiling price list (brand → generic). Procedures are matched against CGHS rates. Items are checked against the IRDAI non-payable list when insurance is involved.
4
We run thirteen deterministic rules
Pure Python. No AI judgment. NPPA ceiling breaches, IRDAI non-payable items, glove and syringe overuse, duplicate lines, room-day mismatches, totals arithmetic — each rule cites the exact notification it tested against.
5
You get a plain-English report
Each flag explains what was found and what you can do about it, with the verbatim legal source attached. Paid tier generates a dispute letter you can send to the hospital billing desk.

Three principles behind the engine

We will reject a feature that violates any of these. The bar for medical billing data has to be higher than the bar for a normal consumer tool.

Hallucination firewall

The AI is allowed to read the bill and narrate findings, but it is never allowed to cite a rate, statute, or regulation that is not in our reference data. Every citation in your report comes from a government notification we have ingested verbatim.

Deterministic where it matters

Every rule that triggers a flag is plain Python tested against fixture bills. If a flag fires for the same bill twice, you get the same answer twice. AI is only used at the read and narrate boundaries.

Public sources only

CGHS rate schedules, NPPA ceiling price notifications, IRDAI Master Circular, DPCO 2013 — every dataset behind a flag is published by an Indian government body and links out from the report.

Common questions about the method

How accurate is the bill reading?+

High-quality PDF bills are read almost perfectly. Photographed bills depend on photo quality — we run held-out benchmark bills before every release and only ship when our match rate stays at 100% on the synthetic baseline. Low-confidence lines are surfaced as 'verify manually' and never appear in dispute documents.

Why thirteen rules and not more?+

Each rule is something a regulator has actually written down — NPPA ceiling under DPCO 2013, IRDAI Master Circular non-payable items, CGHS rate benchmarks. We do not add a rule unless there is a public source behind it. Future verticals (insurance policies, rental agreements) will have their own rule sets.

Does the AI ever invent a citation?+

No. The AI is given only the legal sources that come attached to fired rules. A prompt instruction explicitly forbids introducing a citation, number, or rate not in the data. Low-confidence units are excluded from dispute documents so the AI is never asked to defend uncertain extraction.

What happens to my uploaded bill?+

Original files are stored encrypted on Cloudflare R2 for thirty days so you can re-download the report, then permanently deleted by a scheduled job. PII (patient names) is scrubbed from the structured data before it is kept long-term; hospital names and itemised charges are retained anonymously.

Can the engine handle non-medical documents?+

The engine is built vertical-agnostic — a 'rules' layer plus reference data per vertical. Medical bills are the first vertical. Insurance policies, rental agreements, and salary slips are planned as configuration on the same pipeline, not new code.

Next read