PM-JAY Claim Adjudication

Problem

PM-JAY hospitals submit large, messy claim packets-blurry scans, Hindi/English mixes, handwriting, and 20+ document types where pre- vs post-treatment reports look alike. Manual STG review cannot scale; wrong classification or missed negations (e.g. "no pallor") directly affect reimbursement decisions.

Approach

End-to-end pipeline: rasterize PDFs, two cached Gemma 3 12B calls per page (structure + package clinical flags), deterministic keyword layer with 40-char negation lookback, then intersection fusion (keyword AND VLM unless numeric vitals prove otherwise). Package-specific STG engines emit strict-schema JSON per page plus Pass/Conditional/Fail with provenance.

At a glance

Tech decisions

• Intersection (kw ∧ gem) over pure VLM

  Strict-label evaluation punishes false positives; empirically +3.4 clinical F1 vs Gemma-only override.

• Dual VLM pass with disk cache

  Separates doc taxonomy from clinical fields; reruns cost zero tokens under the 25.5M budget.

• Filename + multi-page rank hints

  Cheap signals recovered mandatory-doc F1 the VLM missed on Indian hospital naming conventions.

• Per-package schema strip

  HI.txt rejects any key-order or link-field mismatch-output is compliance-critical, not approximate JSON.