Helpers on call
B2B journey
KYC at scale.
A life insurer · 200,000 policy applications a month · IRDAI-licensed onboarding
An insurer wants to run KYC at scale without paying a human reviewer for every easy case — and without an off-the-shelf OCR that bills per page.
Marketing claimpii-redaction-frame-ocr
The story · 4 scenes
Beginning to end.
A life insurer · 200,000 policy applications a month · IRDAI-licensed onboarding
·An insurer wants to run KYC at scale without paying a human reviewer for every easy case — and without an off-the-shelf OCR that bills per page.- Scene 01 · Receive a document bundleThe insurer POSTs a multipart bundle: PAN card image, Aadhaar XML, address proof scan, ID selfie. The route accepts every Tier 1–5 ingestion format.
- Scene 02 · Extract under the UIDAI maskMukodi runs VLM extraction with the 8-digit Aadhaar mask applied at the parquet boundary. PAN and 12-digit identifiers are PII-redacted before any model sees them downstream.
- Scene 03 · Flag forged and altered scansThe forgery detector runs PCA-PSI on the embedding space and a confidence band on the classifier. The four common patterns of forgery — copy-paste, font swap, photo-of-screen, repeated stamps — are flagged with reason codes.
- Scene 04 · Return a verifiable bundleThe insurer's API caller receives a structured bundle: extracted fields, masked artefacts, forgery flags, confidence per field. Every artefact is Courtroom-signed.
Helpers on call
The stack · helpers + models
Who's actually doing the work.
Models invoked
Honest limits
What this flow doesn't do.
- OCR confidence falls on very low-resolution scans (< 200 dpi); the API returns the confidence band so the partner can route low-confidence items to human review.
- Forgery flags are advisory — the underwriter still makes the underwriting call under their IRDAI-licensed policy.
- PII redaction is enforced at the parquet boundary; partners receive the masked artefacts and the originals are deleted within the configured retention window.