ABDM gateway · operational Azure India · no egress p50 312ms

Agentic RAG · built for India's health stack

The health router that shows its work.

Setu AI answers clinical questions across your EHR, HIS, labs and claims — it verifies consent, redacts PII, and returns a grounded, cited answer. Every step traced. Nothing leaves Azure India.

Live trace below — watch a real clinical query resolve.
router-v4 · session sess_4f81a trace 4f81a‑2902
M

Show me the full treatment & drug history for Aarav Sharma, ABHA 14‑3344‑2902‑1180. Flag any drug interactions and current adherence.

Parse intent · extract entities
NER · ABHA + name · 40ms
Verify consent artefact
ABDM HIE‑CM · 62ms
Route to source indices
5 candidates · 84ms
Retrieve & rank
k=24 · rerank top‑8 · 106ms
PII guardrail · purpose check
DPDP §8 · 128ms
Compose grounded answer
streaming · 150ms
Treatment & drug history · Aarav Sharmagrounded · 7 sources · 1.84s

54M, primary care at Apollo Andheri. Active: T2DM, hypertension, dyslipidemia1. On Metformin SR, Telmisartan+Amlodipine, Atorvastatin and Empagliflozin2. NSTEMI Jul‑2024, treated medically; LVEF 48%3.

⚠ Atorvastatin + Clarithromycin — myopathy risk ✓ all retrieval served in‑region

The problem

Health data is everywhere.
Trust is the hard part.

One patient's story lives across a dozen systems — ABDM, hospital HIS, lab LIS, imaging PACS, TPA claims. General-purpose AI will answer from any of it: no consent check, no citation, no record of where the answer came from. In a clinical setting, that isn't good enough.

FHIR R4 · EHRHIS · encountersLIS · labs PACS · imagingPMJAY · claimsTPA · denials ICMR · guidelinesABDM · consent

How it answers

One question. Every source. One cited answer.

Setu routes a single question across every system you run — then returns one grounded answer you can check, line by line.

ONE QUESTION Ask once. हिन्दी · English · தமிழ் · मराठी Records ABHA-linked Insurance claims PMJAY · TPA Clinical guidelines ICMR · formulary Uploaded documents PDF · scans · tables Knowledge graph entities linked Structured data computed analytics AGENTIC ROUTER ONE CITED ANSWER Grounded. Consistent across sources. 1 2 3 ✓ faithfulness checked

How it works

Every query takes the same six steps.

The router never improvises. Each question is parsed, consent-checked, routed, retrieved, guarded and grounded — in roughly a second, with a signed trace at the end.

01

Parse intent

Extract entities — ABHA, name, the actual question — before anything is retrieved.

40ms · NER
02

Verify consent

Check the ABDM HIE‑CM consent artefact against the request scope. No valid consent, no data.

62ms · CM gateway
03

Route to sources

router‑v4 scores your indices and picks the top candidates for this exact intent.

84ms · 5 candidates
04

Retrieve & rank

Hybrid search across FHIR, HIS and Rx; rerank down to the eight most relevant chunks.

106ms · rerank top‑8
05

Guardrail & purpose

Redact 23 PII entity types and enforce the declared DPDP §8 purpose before the LLM sees a thing.

128ms · DPDP §8
06

Compose & ground

Stream a cited answer. Anything below grounding score 0.62 is never rendered.

150ms · streaming

Grounded, not guessed

If it isn't in your sources, Setu won't say it.

Every claim links back to where it came from — a FHIR resource, an HIS encounter, a cathlab PDF, an ICMR guideline. Below a grounding score of 0.62, it doesn't render at all. That's how a clinician knows the answer is real, not plausible.

See grounding in a live trace
1
FHIR R4 · Condition
Conditions bundle — 3 active diagnoses
0.94
2
Lilavati HIS · encounters
NSTEMI · coronary angiography, 2024
0.88
3
FHIR · Observation
Echocardiogram · LVEF 48%
0.81
4
Lilavati cathlab · PDF
LAD 70%, LCX 30%, RCA clean
0.76
6
ICMR formulary · 2024
CYP3A4 macrolide interactions · §4.2
0.72

Built for India

Native to the country's health stack.

Setu speaks the systems and rules Indian healthcare already runs on — and keeps data in the country.

ABHA · patient linkage PMJAY · scheme context Azure · India region DPDP-aware · consent posture

One grounded answer — in 15+ languages, Indian and international

हिन्दीEnglishमराठीবাংলাதமிழ்తెలుగుಕನ್ನಡമലയാളംଓଡ଼ିଆभोजपुरी日本語한국어FrançaisEspañolNederlands

Compliance posture

Compliant by construction, not by promise.

DPDP, ABDM and CERT‑In aren't a checklist bolted on afterward — they're how each query is shaped. Here's what's passing, with the controls behind it.

§8
passing

DPDP Act 2023

Every query carries a declared purpose mapped to the §8 schedule.

  • Lawful purpose enforced
  • Granular ABDM consent
  • 23 PII types redacted
M4
certified

ABDM / NDHM

HIE‑CM certified on FHIR R4, with the milestone tracker green through M4.

  • HFR + HPR linkage
  • ABHA + consent
  • Care‑context share
6h
0 open

CERT‑In

Six-hour incident reporting SLA wired to a SOAR bridge. Zero open incidents.

  • Last drill passed
  • NTP sync ±4ms
  • JIT reporting hooks
IN
passing

Data residency

All retrieval served from Azure India. No cross-border egress, ever.

  • AES‑256 at rest
  • TLS 1.3 in transit
  • MeitY / STQC CSP
7y WORM

Audit & trace

Every query trace is cryptographically signed and written to a WORM bucket.

  • Read-only replay
  • Signed lineage
  • 7‑year retention
≤30s

Consent revocation

Consent pulled via ABDM is honoured across indices in under thirty seconds.

  • Purpose limitation
  • Data minimisation
  • Storage TTL 180d

Connectors · across your stack

Connects to the systems you already run.

FHIR-normalised, ACL-aware, change-data-capture where it counts — across EHR, hospital systems, labs, imaging, claims and public health. Add any system with an API.

EHRABDM · HIE‑CMlive · 2s ago
HISApollo Andhericdc · 12s
HISLilavaticdc · 4s
LISMetropoliswebhook
PACSVue · imaginglag 4m
ClaimsPMJAY · TMShourly
ClaimsBajaj Allianzhourly
ClaimsStar Healthlag 22m
NetworkNIN · providersdaily
KnowledgeICMR · guidesweekly
KnowledgeMoHFW · SOPsweekly
PublicIDSP · bulletinsfailed 3h

One router, three surfaces

Built for the people who run the floor.

Clinician

Doctor

Treatment history, drug interactions and adherence — scoped to their own patients, grounded in real records, ready at the point of care.

Operations

Admin

OPD load forecasts, claim-denial root cause, network coverage — the operational questions that keep a hospital moving.

Compliance

DPO

Live DPDP posture, consent revocations, data-principal requests and a signed audit ledger — proof, not assurances.

847k
FHIR records
indexed
4.21M
vector chunks
searchable
312ms
router p50
latency
99.4%
answers
grounded ≥0.62
0
cross-border
egress events

FAQ

Straight answers for buyers.

The questions a hospital, insurer or health-tech team actually asks.

Does it store our patient data?

Setu is designed to run in your environment on Azure, India region, and was built and demonstrated on synthetic / de-identified data — never real patient PHI.

How do we know an answer is right?

Every claim is cited to its source passage and scored for faithfulness. When the evidence is thin, Setu flags the answer or defers to a clinician instead of guessing.

Can it handle our messy files?

Yes — scanned PDFs, borderless and nested-header tables, charts and prescriptions, across 26 formats.

What about counting and reporting questions?

"How many were denied? Which department is highest?" runs through a structured-analytics lane that computes exact figures over your tabular data — not a language-model estimate.

Which languages does it support?

Hindi and many Indian languages, plus English and other international languages — the same grounded answer, in the user's words.

Is it a finished product?

Setu AI is in build and demonstrated on synthetic data; we're setting up pilots. Talk to us about yours.

Bring AI to your health data — safely

See Setu route a real query on your stack.

A 30-minute walkthrough on your sources, your roles, your compliance posture. Built in India, for India's health stack.

or email kumar.j5098@zohomail.in