BillGaurd - Cause-Aware Cross-Layer Verification of Indian Financial Documents

BillGaurd is a cause-aware, cross-layer evidence-review framework for Indian financial documents. It treats a mismatch as an observation to investigate rather than an automatic fraud verdict: the discrepancy may come from recognition error, an inconsistency in the issued source, a later edit, or unavailable evidence.

The research implementation preserves the submitted artifact and SHA-256 provenance, routes native PDFs, rendered images, scans, and screenshots according to the evidence they retain, gathers OCR, visual, PDF-structural, financial, GST/GSTIN, HSN/SAC, and QR observations, and aligns them to canonical fields for reviewer-oriented reporting.

  • Cause-aware field outcomes - Reports each material field as Verified, Likely Analysis Error, Source/Document Inconsistency, Probable Tampering, or Unresolved/Abstain.
  • Representation-aware acquisition - Keeps native PDF text, rendered OCR, image, scan-like, and screenshot evidence distinct instead of silently treating every representation as equivalent.
  • Reversible OCR modes - Supports Indra-only, RapidOCR, and combined operation; materially different readings remain support-only evidence rather than automatic tampering claims.
  • Evidence-aware review - Separates an unavailable check from a check that found no anomaly, keeps a document-level structural model advisory, and abstains when extraction or field alignment is insufficient.
  • Indian financial validation - Combines arithmetic, GST/GSTIN, HSN/SAC, QR, PDF, visual, and provenance observations while keeping offline fixtures and optional semantic assistance bounded.

The research and review papers describe an engineering study, not a production fraud detector or legal-authenticity claim. Their central boundary is deliberate: OCR quality, document review cues, field attribution, and cause decisions are reported as different tasks rather than merged into one score.

Technical approach

BillGaurd uses a layered acquisition and evidence-alignment pipeline with PyMuPDF/OpenCV processing, Indra and RapidOCR adapters, deterministic arithmetic and regulatory checks, optional bounded Nemotron/NVIDIA NIM assistance, and guarded single-field reconstruction. The direct cause engine reached 1.0000 accuracy and macro-F1 on 20 held-out synthetic cases, but the complete 120-run evidence-recovery pipeline reached only 0.0167 coverage and abstained on 0.9833 of runs. This exposes extraction and field alignment as the main bottleneck rather than hiding it behind a strong isolated score.

On 46 held-out documents from 23 untouched pairs, the advisory PDF-structural model reached 91.30% accuracy and 0.9556 ROC-AUC; separate calibration reduced Brier score from 0.0641 to 0.0454. On the same first 12 FUNSD pages, Indra-only produced CER/WER 0.3461/0.5310 in 9.06 seconds versus RapidOCR at 0.3794/0.5631 in 41.32 seconds. These are bounded OCR-reading and document-review results, not field-cause or universal invoice-forgery accuracy. Read the structured review and publication details on the Publications page.