01 Text detection DL
task Locate lines of text on the page input Page image model Segmentation-based detector output Bounding boxes / polygons around text areas
used by VERIDOC
01 Intro
Claim documents are read, checked for authenticity, and cross-referenced with related claims using machine learning models.
02 Pipeline
Four stages
02 Text recognition DL
task Convert text image crops into characters, including handwriting input Text line image crop model Sequence recognizer output Recognized text
used by VERIDOC
03 Forgery signal SIGNAL
task Flag image areas that may have been re-edited input Document image model Error Level Analysis. JPEG compression differences turned into a per-area anomaly map output Per-area anomaly map
used by VERIDOC
04 Risk scoring
task Combine all signals into one explainable risk score input Signals from the detection, recognition, and forensic stages model Combines signals from the other stages into one explainable score output Risk score together with the signals behind it
used by VERIDOC
03 Systems
VERIDOC
Claim document forensics · Developed in-house by our internal team
Tampered claim documents are hard to spot by eye. Small edits to numbers, dates, or stamps often slip past manual review.
Photo or scan of a document, PDF → Structured text, a map of suspicious areas, a risk score with its reasoning
Relata
Document relationship & semantic search
Related claims live across different documents and different systems finding them by keyword alone misses cases that describe the same thing in different words.
Recognized text from VERIDOC → Related documents and cross-referenced claims, ranked by meaning
04 Principles
- 01 Developed in-house by our own team, trained specifically to catch fraud.
- 02 Every score comes with a reason you can trace.
- 03 The model learns from real cases and improves through the team's feedback.
05 Contact
For technical questions about VERIDOC or Relata, contact it@tbmperkasa.net.