Evidence network
How NRD works
NRD is the structured-evidence layer between a claim and a decision. Every step is auditable. Nothing is inferred without a source.
The evidence network, step by step
Seven stages, fully logged. Each stage either produces a verifiable output or hands off to a human before proceeding.
- 01
Intake
Records arrive via WhatsApp, field-exec upload, email or insurer feed. Each file is timestamped, assigned a unique reference and written to an immutable S3 prefix in ap-south-1.
- 02
OCR — with an honest gate
Printed and typed documents pass through OCR with a per-page confidence score. Vernacular text (Hindi, Marathi, regional scripts) and handwritten content trigger a confidence threshold check. Documents below the threshold are flagged for a human reviewer before extraction continues. We do not silently pass low-confidence output downstream.
- 03
Entity extraction
From confirmed OCR text, NRD extracts typed fields: medical diagnoses, ICD-10 codes, treating doctor, hospital and dates, itemised charges, witness names and statements, and policy / FIR references.
- 04
Quality control
Thirteen deterministic checks run on every record set: page count vs declared total, legibility score per page, PAN/Aadhaar field detection, duplicate document hash, date-sequence consistency, and cross-document contradiction flags.
- 05
Gap analysis
NRD compares each claim's evidence set against a document checklist for its claim type (head injury, fracture, soft tissue, death). It returns a structured gap list — missing medicals, missing FIR, absent discharge summary — that a surveyor or panel lawyer can act on.
- 06
Summary
A structured evidence summary is assembled: provider timeline, billed vs paid totals, reserve recommendation based on extracted figures, and causation narrative sourced from the clinical record.
- 07
Chain-of-custody certificate
Every document in the final claim file is SHA-256 hashed, timestamped to IST and locked with S3 Object Lock (WORM). The resulting certificate is court-ready under the Bharatiya Sakshya Adhiniyam, Section 63.
Accuracy architecture
We group every output by how it is produced and what weight you can give it. Honest labelling is the point.
| Tier | How produced | Examples | Weight |
|---|---|---|---|
| Deterministic | Pure computation from extracted structured fields. No model involved. | MACT compensation (Pranay Sethi formula), court fees, limitation period, date-sequence checks, document hash verification. | Full trust — auditable formula, reproducible. |
| Extraction-gated | Derived from OCR/NLP extraction, which is confidence-scored. If confidence is below threshold, a human reviews the source document before the field is used. | ICD-10 codes, treatment dates, itemised charges, provider names, vernacular-script text, handwritten content. | Verified extraction — traceability to source page. Vernacular / handwritten output always carries a human-in-loop flag. |
| Predictive | Model-generated signals that assist a decision but are not the decision. Not used for binding outputs. Clearly labelled in the UI. | Reserve-adequacy range estimate, fraud-indicator score, defence-propensity signal, settlement ceiling range. | Decision-support only. A claims officer reviews and confirms before action. |
The vernacular-OCR human gate is not a disclaimer — it is how the pipeline is built. Low-confidence documents do not move forward without review.
Glossary
- MACT
- Motor Accident Claims Tribunal — the Indian forum that adjudicates compensation for motor-accident injury and death claims.
- FMF / FMR-1
- IRDAI Fraud Monitoring Framework and its FMR-1 return — mandatory fraud-monitoring reporting for Indian insurers since 1 April 2026.
- DPDPA 2023
- India's Digital Personal Data Protection Act — governs processing and residency of personal data.
- CERT-In
- India's Computer Emergency Response Team directions — incident reporting and log-retention requirements.
- BSA S.63
- Bharatiya Sakshya Adhiniyam, Section 63 — admissibility of electronic records as evidence.
- Pranay Sethi
- Supreme Court precedent setting the deterministic basis for MACT compensation computation.
See the evidence network on your claims.
We'll scope a pilot on a slice of your motor/MACT book.