Eight agentsone consensus
They do not run in sequence. All eight fire simultaneously, and a Weighted Evidential Consensus combines their outputs. Where they disagree materially, a structured debate round runs before the answer is returned. Move the cursor to look underneath.
Clinical NLP
Reads the doctor’s notes in any of 12 Indian languages and extracts symptoms, negations, severity and urgency. “Bukhar teen din se hai” is the same input as “fever 3 days”.
Hypothesis Generation
Ranks differential diagnoses by Bayesian reasoning, starting from base rates and updating on symptom evidence. Constrained to ICD-11 codes, so it cannot invent a condition.
Lab Correlation
Correlates CBC, LFT, RFT and vitals against each hypothesis. Rule-based where it matters: platelets under 50K always flags dengue, CRP over 100 flags sepsis. Reads lab PDFs by OCR.
Drug Safety
Checks every drug pair for interactions and verifies NLEM availability. Flags NSAIDs in pregnancy and aminoglycosides alongside loop diuretics.
Temporal History
Reads the patient’s full history across past visits and separates a new condition from a relapse or an adverse drug effect.
Evidence Retrieval
Searches PubMed, Cochrane reviews and NMC guidelines per diagnosis, ranking RCT above cohort above case series, and returns citations with the differential.
Imaging Analysis
Analyses uploaded X-rays, CT scouts and ultrasound stills for consolidation, free fluid and mass lesions, and feeds the findings back in as evidence.
Epidemiological Prior
Pulls live IDSP district alerts and adjusts every prior for the patient’s region. Dengue in Sikar, malaria in Alwar, TB everywhere.
The whole hospital, one clinical system.
Case workspace
Case list on the left, the pipeline running in the middle, the differential appearing on the right as agents complete. Accept a diagnosis and it writes an audit event; prescriptions, labs and uploads hang off the same case.
Four voice agents
Patients speak, doctors speak, and the two are rarely speaking the same language.
Patient health timeline
Every interaction writes an event, all of it embedded in Pinecone for semantic search.
Document vault
X-rays, lab reports, clinical photos, old paper records. Tesseract extracts, Pinecone indexes, Doctor Line recalls it by voice.
Reception dashboard
Live Aarogya calls, transcript as the patient speaks, specialty-matched doctor suggestions, one click to create the case.
Multi-tenant by design
Every hospital is its own organisation, every query scoped by organization_id. Roles for hospital_admin, doctor, nurse, front_desk, pharmacist. Invite-only onboarding, zero cross-tenant leakage.
Input. Analyse. Output.
The doctor writes or speaks
Notes in any of 12 Indian languages, dictation through Scribe, lab PDFs, X-rays, or an Aarogya intake call already on file.
Eight agents fire at once
NLP, hypotheses, labs, drug safety, history, evidence, imaging and district epidemiology run in parallel, then Weighted Evidential Consensus.
A ranked differential, under 90 seconds
Confidence, ICD-11 code, supporting and contradicting symptoms, citations and a drug interaction check. Accepting one writes an audit event.
One platform, three very different hospitals.
The full pipeline on a bad connection. Voice input in Hindi, offline capability, no specialist required within 100 km.
Validated, not asserted.
MACS-Dx — Multi-Agent Clinical Supervision for Diagnostics — was evaluated on DDXPlus, the NeurIPS 2022 benchmark: 500 cases across 49 pathologies. Every number here comes from that run.
From the doctors using it
“I see eighty patients a day and the nearest specialist is a hundred kilometres away. I dictate in Hindi, I get a differential before the next patient sits down.”
“Twelve doctors were consulting each other over personal WhatsApp. Now every consult is a case with an audit trail, and the district dengue alert reaches all of them at once.”
“We had turned down three clinical AI tools because none of them spoke FHIR. This one wrote into our existing records on day one, with citations my registrars can check.”
Pricing
Rural primary care never pays. It is the same platform either way.