AI-powered VA claims analysis that discovers claimable conditions from medical records, identifies rating errors in VA decisions, and generates evidence-backed appeal packages.
A local-first, privacy-preserving platform built for veterans navigating the VA disability claims process. The system ingests medical records (STRs, C&P notes, decision letters), discovers claimable conditions using hybrid vector search and LLM analysis, validates findings against 38 CFR Part 4 regulatory criteria, and produces citation-backed reports with specific appeal recommendations.
Medical records are indexed with both dense (BAAI/bge-small-en-v1.5) and sparse (Splade) embeddings, fused via Reciprocal Rank Fusion. This hybrid approach ensures both semantic similarity and exact terminology matching — critical when searching for specific diagnostic codes, medication names, or medical terminology across hundreds of pages.
Every condition is evaluated from two opposing perspectives simultaneously: a Veteran Advocate (seeking the highest defensible rating) and a VA Rater (applying the regulatory standard). A Reconciler then evaluates both positions against actual 38 CFR criteria text retrieved from a dedicated regulatory collection. This mirrors the real VA adjudication process and eliminates blind spots that single-perspective analysis misses.
Rather than relying on a single LLM pass, condition discovery uses a deterministic multi-phase pipeline: source indexing, diagnostic scanning, LLM extraction, normalization, per-condition evidence search, CFR criteria mapping, service connection assessment, condition classification, and cross-condition pattern analysis. Each phase is independently testable and produces structured intermediate results.
| Feature | Technical Implementation | Business Value |
|---|---|---|
| Condition Discovery | 9-phase pipeline with hybrid RAG search across medical records | Discovers conditions veterans miss — including presumptive, secondary, and bilateral claims |
| Decision Letter Analysis | Regex-based parser + 6 error categories + 8 C&P inadequacy detectors | Identifies specific rating errors and procedural failures in VA decisions |
| Evidence-Backed Reviews | Per-condition Qdrant search with relevance filtering and citation validation | Every finding is traceable to a specific page and quote in the medical record |
| Rating Projection | VA combined rating math with bilateral factor, TDIU, and back pay calculation | Shows veterans what their rating should be vs. what VA awarded |
| Appeal Routing | Decision matrix mapping error types to optimal appeal lanes (HLR, Supplemental, BVA) | Recommends the fastest path to resolution with specific form references |
| Form Completion | Deterministic mapping of discovered conditions to SHA Part A (DD Form 3146) questions | Pre-fills ~100 self-assessment questions with evidence citations |
| Benefits Briefing | Federal + Colorado state benefits lookup by rating threshold | Immediate clarity on compensation, healthcare, education, and housing benefits |
| Report Generation | Structured markdown to Word document conversion | Professional, citation-backed reports suitable for VSO submission or self-filing |
Screenshots coming soon — the platform includes a React frontend with file upload, condition review cards, rating summary, and benefits dashboard.
| Metric | Value |
|---|---|
| Test suite | 1,324 tests across 27 files |
| Code coverage | 68% |
| Conditions parsed (real data) | 43 from a single 59-page decision letter |
| Discovery pipeline phases | 9 deterministic phases |
| SHA form questions mapped | ~100 (DD Form 3146, all sections) |
| Error categories detected | 6 rating errors + 8 C&P exam inadequacies |
| Reference data | 79 DBQ templates, full 38 CFR Part 4, 18+ VA forms |
| Analysis depth tiers | 3 (quick/standard/deep) |
| API endpoints | 15+ REST endpoints (FastAPI) |
| Layer | Technology | Purpose |
|---|---|---|
| Language | Python 3.12+ | Primary implementation |
| API Framework | FastAPI | REST API with automatic OpenAPI documentation |
| Frontend | React + Vite + TypeScript + Tailwind CSS | Single-page application |
| Vector Database | Qdrant | Hybrid dense+sparse search with collection isolation |
| Embeddings | fastembed (bge-small-en-v1.5 + Splade) | Local dual-vector generation (no API calls) |
| LLM | Claude (Anthropic) | Primary analysis engine with structured output |
| Document Processing | PyMuPDF | PDF text extraction with date scoring |
| Report Generation | python-docx | Markdown to Word document conversion |
| Testing | pytest + mutmut | Unit, integration, and mutation testing |
| CI/CD | GitHub Actions | Automated test, lint, security scan pipeline |
See ARCHITECTURE.md for the full system design with C4 diagrams, data flow, and security posture.
The VA disability claims process is complex, adversarial, and consequential. Veterans separating from service must:
- Discover all conditions that may qualify for VA disability compensation
- Document each condition with medical evidence meeting 38 CFR Part 4 criteria
- File claims with the correct forms, evidence, and legal basis
- Review VA decisions for errors (which occur in ~25% of cases per VA OIG reports)
- Appeal incorrect decisions through the correct lane (HLR, Supplemental, or BVA)
This system automates steps 1-5 with citation-backed evidence analysis, regulatory criteria mapping, and structured output that meets VSO submission standards.
Copyright 2026 TJ Neary. All Rights Reserved.