In detail · September 2026

The pages behind the pitch.

How the model is built, what it reads, how we measure it, how it handles data, who else is here, how we sell, and the team and the round.

1 · How it is built

Open base model, Indian corpus, retrieval, a source under every answer.

Nidaan is not trained from zero. It starts from an open Indian base model, learns Indian healthcare from a corpus written out of the public rules, and answers through a retrieval index so that every answer can show the passage it rests on. Five parts, in the order a question passes through them.

1 · Sources

PM-JAY packages and the adjudication manual · state schemes · IRDAI rules and circulars · policy wordings · STGs, ICMR, NLEM, NPPA · partner questions and outcomes, with consent.

2 · Retrieval index

Every document chunked and indexed. Version-dated, so a new circular supersedes the last. Returns the passage that becomes the citation.

3 · The model

Open Indian base model (Sarvam 30B/105B class, Apache 2.0). Fine-tuned on Indian healthcare Q&A written from the sources. Learns from consented claim outcomes. 22 languages, any script.

4 · Safety layer

Cites the source or says "not sure". Human review of flagged claims. Audit log of every answer. No partner data crosses editions.

5 · Delivery

API, per call · India-cloud edition · on-premise edition · custom edition per payer or scheme.

Compute: fine-tuning and serving on subsidised IndiaAI GPUs and Indian cloud regions; on-premise editions run on the customer's own hardware. Architecture as planned for phase 1; the base model is confirmed after benchmarking candidates.

2 · What it reads

The corpus, and on what terms.

Almost everything a claim is decided on is public and written down. The model reads it all, with the licence and reuse terms of every source confirmed before training; nothing is scraped from behind a login. Partner data is the one private source, and it enters only with consent, de-identified, per partner.

SourceWhat it gives the modelAccess and licence
NHA Health Benefit Package 2022 and PM-JAY adjudication manual1,949 procedures with rates, packages, pre-authorisation and documentation rulesPublic · terms of reuse to confirm with NHA
State health schemes (11+)Each state's packages, rates and rules where they differ from PM-JAYPublic, per state · collected state by state
IRDAI regulations, master circular (29 May 2024), circularsCashless deadlines, claim settlement and grievance rules, data-in-India requirementsPublic
Insurers' policy wordings (100s)Exclusions, waiting periods, sub-limits, room-rent caps, co-pays by productPublic on insurer sites and IRDAI filings · reuse terms to confirm
MoHFW Standard Treatment Guidelines, ICMR guidanceWhat good care looks like for each condition; what a payer may questionPublic
NLEM, NPPA ceiling pricesWhich medicines are essential and what they may costPublic
ICD-10 (WHO), SNOMED CT (NRCeS)Diagnosis and procedure codes the exchange and payers useLicensed · free for Indian use via NRCeS; WHO terms
Bhashini, AI4Bharat, Sarvam modelsIndian language coverage, speech, an open base modelOpen · Apache 2.0 and government terms
Partner questions and claim outcomesWhat people actually ask, and what happened to the claimConsented, de-identified, per partner; DPDP; never shared across editions

3 · How we measure

The Indian claims benchmark: how we know it works.

A model that answers Indian claims questions has to be scored on Indian claims questions. There was no such benchmark, so we are writing one, and we will publish it.

The benchmark

300 real questions across the five stops (admission, treatment, filing, review, payment), written from real cases, in Hindi, Hinglish and English first; more languages as partners bring them.

Expert answers from our two doctors and the billing and TPA experts on the partner side, with the source each answer rests on.

Compared against Claude, GPT, Gemini and Sarvam on the same questions, same day.

Published at month 6; re-run every quarter as the corpus and the model change.

What we score

Correct answer: matches the expert answer.

Correct source: cites the passage the answer rests on.

Knows when to stop: says "not sure" rather than guessing.

Language: the same score whether typed in Devanagari or English letters.

Cost and time per answer at India's volume.

Honest limits. Nidaan supports the person deciding; it does not diagnose, and a human reviews every flagged claim. Every failure is logged, classified and fed back into the next training run. Targets and first results are added at month 6.

4 · Trust, privacy and regulation

Built for India's rules on data, not around them.

Health data in India is now sensitive personal data by law, insurers must keep primary data in India, and the claims exchange has a standard format. Nidaan is designed to those rules rather than routed around them.

Data protection

DPDP Rules notified 14 November 2025; health data is sensitive personal data; penalties up to ₹250 crore; full force by May 2027. Consent recorded per partner; de-identified before training; an India-based Data Protection Officer when we cross the thresholds.

Insurance rules

IRDAI: insurers keep primary data in India; the master circular sets the 1-hour and 3-hour cashless clocks and settlement timelines. Our payer editions run in India or on the insurer's own servers; nothing leaves.

Standards

ABDM: ABHA identity and consent. NHCX: FHIR-based claim formats between hospital and payer. Nidaan reads and writes those formats, so partner apps plug in without a new standard.

Positioning

Decision support: a human decides; no diagnosis; not a medical device in phase 1. The clinical layer planned for year 2 is assessed against CDSCO software-as-medical-device rules before launch.

In one line: a source under every answer, an audit log, on-premise editions, consent and de-identification by design, and no partner's data ever used for another partner.

5 · The landscape

Who else is here, and why none of them is this.

The global labs reason brilliantly and now run inside India; India's own models are open and speak 22 languages; Indian health AI is real. None of them knows PM-JAY, IRDAI or a policy wording, and none learns from what happens to an Indian claim. We build on the open models rather than compete with them, and add the layer none of them has.

WhoWhat they doWhere Nidaan sits
Global labs' health products
Anthropic (Claude for Healthcare), OpenAI (OpenAI for Healthcare, ChatGPT Health), Google (MedLM, HAI-DEF)
Strong general and clinical reasoning; health tools built on US systems (CMS, NPI, US records); closed weights; in-country inference in India from 2026Not our base and not our layer: they lack Indian rules, wordings and outcomes. We build on open models and add the Indian layer.
India's general models
Sarvam (30B, 105B; Apache 2.0), BharatGen (Param2), Krutrim
22 languages, voice, open weights, subsidised compute; general purposeOur base model. A supplier, not a competitor.
Indian health AI
Fractal Vaidya 2.0, Eka Care
Clinical reasoning for citizens and doctors; scribing and record reading; no public claims or scheme coverageAdjacent. Natural builders on Nidaan for the claims and scheme layer.
Claims software and TPAs
NHCX itself; TPA in-house tools; claims automation startups (Vitraya, Consint.AI, ClaimBuddy; RapidClaims and RISA Labs for the US)
Workflow, OCR and rule engines; each teaches a general model its own rulesBuilders on Nidaan, not rivals: they own the app, we own the model. Some may compete with our payer editions.
Global analogues
Hippocratic AI, Abridge, Anterior, Cohere Health
Domain models and agents for US health systems and payers; safety-first positioning; large roundsProof that a domain-model company is a category. None is built for India.

As of September 2026, from public announcements, funding news and company sites. Frontier models lost up to 24 points on Indian languages typed in English letters in the Script Gap study (arXiv, December 2025), and Indian text needs up to 8× the tokens on GPT-4's tokeniser (The Tokenizer Tax, arXiv, July 2026).

6 · How we sell

Free for design partners, then builders pay.

We sell the model, not apps, in three ways: per call through the API, as a yearly licence for a private Nidaan, and as a custom edition tuned on one payer's own decisions. The first five design partners use it free for twelve months, on purpose: their questions and consented outcomes are worth more than the fee.

Pay per call · API

₹20 a claim check, a small fraction of the roughly ₹500 it costs to process a claim today (NATHEALTH); questions priced per thousand. Free for the first five design partners.

Enterprise licence · yearly

A private Nidaan in an India-hosted cloud or on-premise, for hospital chains, insurers, TPAs and hospital software vendors. Nothing leaves your walls.

Custom editions · build plus licence

Tuned on one insurer's, TPA's or PM-JAY state agency's own policies and past decisions. Their data stays theirs and is never used for anyone else.

Who holds the budget line

33 insurers writing health (7 standalone, 26 general) · the TPAs that handle 69% of health claims, about 2.25 crore a year · 33 state health agencies running PM-JAY and their own schemes · 39,000+ PM-JAY hospitals, reached through their software vendors · 118 crore people covered: 60 crore+ under PM-JAY, 58 crore privately insured.

India decides about six crore claims a year, ₹1.45 lakh crore of them, and spends about ₹3,000 crore deciding them. That is the budget line we sell into; India's healthcare AI market is forecast at ₹45,800 crore by 2034 (IMARC).

The five design-partner slots

One each: a hospital software vendor, a hospital billing desk, a TPA, an insurer and a patient claims app. Free for twelve months in return for questions and consented outcomes; pilots in Noida; founder-led. Prices for licences and editions are set with the first ten buyers.

Apply by email →

7 · The team and the round

Six co-founders, one model, ₹12.5 crore for the first 15 months.

Six co-founders, two doctors and four technologists, 10 to 16 years each in healthcare: US revenue-cycle specialists who built 10+ AI platforms for large American healthcare companies, each earning multi-million-dollar recurring subscription revenue. Entity details, cap table and founder profiles are shared in the data room.

The raise

₹12.5 crore (about $1.3M) for the first 15 months: the 12-month plan plus a 3-month buffer. 70% team, 15% compute and Indian corpus, 5% design partners and sales, 10% compliance, operations and buffer. ₹2.5 crore committed, 20% of the round. Instrument and pre-money are set with the lead.

What it proves by month 15

Nidaan beats the general models on 300 real Indian claims questions · five design partners live on the API and first paying customers signed · first payer edition in pilot with one TPA or state scheme. These are the milestones for the next round.

Where we are today

nidaan.ai live · the Indian source map done · the 300-question benchmark designed, first run against the general models next · five design-partner slots open · nothing here assumes a signed contract. By 2028: the model every Indian claim is checked against.

Non-dilutive money that stretches the runway, none of it assumed: IndiaAI Compute Portal, MeitY SAMRIDH, UP Startup Policy 2026 (as reported), NHCX/ABDM incentives, BIRAC BIG and the RDI Fund, cloud and model credits. Terms are shared in the data room.

8 · Sources

Where every number comes from.

All figures as of September 2026; conversions at ₹96 per US dollar. Links on request.

The problemIRDAI 2023-24 claims data and 2024-25 annual report; IRDAI master circular (29 May 2024); NHA Health Benefit Package 2022; National Health Survey 2025 via ORF; Business Today (March 2025); AccredReady; Business Standard (claims disallowed, repudiated, pending).
Why nowIRDAI (29 May 2024); NHA, NHCX first claim (June 2024); MeitY DPDP Rules (14 November 2025) and India Briefing timeline; India AI Impact Summit and IndiaAI Mission (February 2026); Sarvam (March 2026); Anthropic and AWS in-country inference (3 August 2026).
Why a general model won't do itPublic announcements by Anthropic, OpenAI, Sarvam, BharatGen, Fractal, Eka Care; the Script Gap study (arXiv, December 2025); The Tokenizer Tax (arXiv, July 2026); Health Minister statement on PM-JAY coverage (23 September 2026).
MarketIRDAI Annual Report 2024-25 and IRDAI data via Business Standard; Government statements on PM-JAY (March 2025, August 2026); NATHEALTH (processing cost per claim); IMARC Group healthcare AI estimate.
Business model and moneyNATHEALTH note on NHCX (FY24: 3.26 crore claims, average claim ₹31,086); IndiaAI Mission; MeitY SAMRIDH, UP Startup Policy 2026 (as reported), BIRAC, TDB; cloud credit programmes.

Talk to us

Investors, partners, press.

contact@nidaan.ai · Noida. The deck, the data room and the benchmark results at month 6 are available on request.

Pilots open in Noida

Email contact@nidaan.ai

Back to the overview