Identity and PAN check. Used in: Loan application. 1 page. Specimen drawn to the real layout, for a fictitious person and bank; the emblem, logos and real QR data are left out.
- 1. Classify
- 2. Read fields
- 3. Check formats
- 4. Cross-check
- 5. Route
- 6. Decide
Review queue
Fields under the 80% bar after routing, or named by a failed check. Each decision is saved as eval and training data.
Run a document to fill the queue.
This batch
Documents with no manual touch
–
Field accuracy vs answer key
–
Reviewer decisions saved
0
Cost vs all-large
–
Measured live from the runs above, against the answer key we wrote for these 5 documents. Large-model price assumed at $0.012 a page.
Schemas first
Each document type has a field schema; the model fills it and the schema rejects anything malformed.
Checks a model can't skip
PAN and IFSC formats, real dates, totals that add up, names that match across documents.
Routing on cost
A small vision model on Dvar first, a large model only for the hard page, cost tracked per document.
Review that teaches
Low-confidence or failed fields go to a person; their corrections become eval and training data.
Traces and evals
Every run traced in Dvar Evals with an eval set per document type, so a prompt or model change is tested first.
Data handling
PII masked in logs, India hosting, retention rules that fit DPDP.
Written for the demo: the 5 specimen documents (drawn to the real PAN, masked Aadhaar, CTS-2010 cheque, bank statement and salary slip layouts, for a fictitious person and bank) and what each model read from them, with its confidence. Real code running here: the schemas, the checks, the routing rule, the review queue and the scoring. The next step is to run this harness on a redacted sample of KreditBee's own documents.