AI verification engine
The engine that reads, reconciles and scores every document.
Papertrail's AI handles the reading and reconciling, not the deciding. Documents are extracted, cross-checked against source and scored for tampering — and every verdict a client sees has cleared a human reviewer, not just a model.
- 20+sources cross-checked per case
- <2 mintypical extraction time per document
- 0verdicts published on a machine guess alone
Extracted fields
- Employee name0.00
Priya Ramanathan
- Employee ID0.00
IN-0447192
- Designation0.00
Senior Analyst
- Date of joining0.00
11 Jul 2019
- Date of relieving0.00
28 Feb 2023
heldDigits overlap the office stamp — sent to a reviewer with the scan open.
- Reason for exit0.00
Resigned
- Signatory0.00
R. Menon, HR Manager
0 auto-accepted · 0 held for a person
A field below threshold waits. It is never guessed.
The engine extracts seven fields from a relieving letter with a confidence score on each. Six clear the threshold; the date of relieving scores 0.61 because the office stamp overlaps the digits, so it is held for a human reviewer instead of being published.
Document intelligence, fraud signals, adjudication.
Each layer does one job. Together they turn a folder of uploaded PDFs into a case file with a defensible answer in it.
OCR extracts every printed field — names, dates, designations, serial numbers — and structures them for comparison. Layout-aware parsing means a table on a payslip and a paragraph in a relieving letter are both read correctly, not just scanned for keywords.
Font consistency, metadata and edit history, re-compression artefacts from a second photograph of a photograph, and letterheads already seen on an unrelated candidate's file — each produces its own signal, scored against a threshold rather than a single pass/fail guess.
A one-line title discrepancy and a fabricated employer are not the same severity, and the engine doesn't treat them as one. Findings are ranked, genuine disagreements are separated from harmless variation in wording, and anything above a confidence threshold routes to a reviewer before it reaches your report.
One relieving letter, four stages.
This is the same document moving through all three layers above. Step through it — OCR extraction, field validation, fraud-signal scoring, verdict.
Relieving letter
REF/NR/2024/1187 · 18 Feb 2024
- Employee name
- Ananya Iyer
- Designation
- Senior Associate
- Date of joining
- 03 Jun 2021
- Date of leaving
- 14 Feb 2024
- Reason for exit
- Resigned — eligible for rehire
The scanned letter is read field by field — name, designation, dates, exit reason — and structured into data before anyone opens the case. No manual re-keying, and nothing is inferred that isn't printed on the page.
5 of 5 fields extracted
Touring all four stages — pick one to hold it
What a human reviews before a flag reaches a client.
A flag the engine raises does not go to your report. It goes to a queue. This is that queue — the evidence a reviewer has in front of them, the three calls they can make, and the line each call writes to the case log.
Nothing in this queue is on your report yet. It is on somebody’s screen.
Relieving letter font changes mid-line on the date of exit
On the reviewer’s screen
- The uploaded PDF at full resolution, with the two glyph sets the engine separated highlighted side by side.
- The file's metadata: created in a PDF editor 3 days after the stated relieving date, producer string rewritten.
- The same employer's letterhead as it appeared on four earlier, unrelated cases — for comparison, not as proof.
The call they can make
The finding is written into the report as a document-integrity discrepancy, with the reviewer's own annotation and the EPFO record attached as the corroborating source. The candidate is notified and given the dispute window.
14:22 IST · A. Nair confirmed tamper signal 0.71 · evidence: EPFO/UAN · candidate notified
The signal is recorded as reviewed and not carried forward. Nothing about it appears on the client's report — a false positive stays an internal event, not a mark against a candidate.
14:22 IST · A. Nair dismissed tamper signal 0.71 · reason: re-scan artefact · not published
The case moves to direct employer confirmation before anything is published. The report shows the check as in progress with the reason, rather than showing a finding the reviewer isn't sure of.
14:22 IST · A. Nair escalated to employer HR · case held open · client status: in progress
Pick a call to see what it does to the case — and what it writes to the log.
Hover or focus a figure to see what sits behind it
The AI engine, answered
The engine reads documents, cross-checks values and scores fraud signals. It does not decide whether a discrepancy disqualifies a candidate — that stays with your hiring team, after a human analyst has reviewed anything the engine flagged.
Discrepancy is modelled explicitly, not inferred from whether two strings match. "Resigned" and "Resigned · eligible for rehire" are treated as corroboration; a designation that's a full level different is treated as a finding. Every comparison is field-by-field, not a blunt text diff.
The fraud-signal layer scores things like font inconsistency, stripped metadata and letterhead reuse. A score above threshold routes the actual document to a reviewer trained to examine it — the case does not auto-fail, and it does not auto-pass either, until a person has looked at it.
Yes, which is why nothing it produces reaches a client-facing verdict unreviewed. Confidence thresholds, independent-source checks and mandatory human review on anything flagged exist specifically because an automated read is a proposal, not a conclusion.
See the engine on a real case
Bring us a document, we'll show you the read.
Send a sample relieving letter or certificate and we'll walk your team through exactly what the engine extracts, flags and escalates.