Rybera

Denial defense infrastructure for commercial health insurers and TPAs.

Rybera intercepts fraudulent and miscoded claims before payment — scoring each one in real time, surfacing anomalies to Special Investigations Unit analysts, and generating court-defensible evidence packs for recovery and appeal.

---

The Problem

Commercial health insurance fraud costs the US system an estimated $36–50 billion per year (NHCAA: ~3–5% of ~$1T commercial premium). The dominant fraud vector today is AI-scribe-driven upcoding: providers using AI documentation tools that systematically inflate E/M codes, creating a statistically detectable but individually plausible pattern that legacy rule engines miss entirely.

Rybera catches it.

---

Product Architecture

Layered Fraud Scoring Pipeline (real-time, < 2s p99)

POST /v1/claims/score
        │
        ├── Ingest & normalize (patient_id hashed at boundary)
        ├── NPPES / OIG LEIE enrichment
        │
        ├── Layer 2 — AI Inflation Detection
        │     Clinical note embedding · coherence scoring ·
        │     speed-of-documentation signal · condition delta
        │
        ├── Layer 3 — Provider Pattern Analysis
        │     24-month rolling profile · E/M distribution z-score ·
        │     AI-scribe z-score · billed amount z-score · changepoint detection
        │
        ├── Layer 4 — Policyholder Pattern Analysis
        │     Cross-provider claim velocity · distinct provider count ·
        │     escalation rate · high-E/M concentration
        │
        └── Composite Score (0–100) + Routing Recommendation
              PAY  →  normal adjudication
              HOLD →  pend for SIU review
              ESCALATE → immediate SIU investigation

Clinical Validation Pipeline (E/M coding integrity)

validation_graph.py — LangGraph validation pipeline
        │
        ├── Clinical Validation Agent   (2021 AMA MDM scoring, forced tool-use)
        ├── AI-Scribe Detection         (HealthScribe / DeepScribe / Nabla signatures)
        ├── NCCI / MUE Code Edits       (deterministic bundling + volume checks)
        ├── Modifier Abuse Checks       (deterministic — fraud.py primitives)
        └── Disposition Synthesis       (fraud score + recommendation)

Validates E/M coding against the documented clinical note, catches NCCI bundling violations, modifier misuse, and AI-scribe-inflated documentation.

POST /api/v1/validate-claim           Full LangGraph validation pipeline
POST /api/v1/validate-claim/batch     Bulk validation
POST /api/v1/validate-claim/stream    SSE stream — watch agents fire live
GET  /api/v1/validations              Paginated validation history
GET  /api/v1/validations/{id}/audit-trail  Immutable SHA-256-chained audit log

SIU Portal (case management)

All SIU endpoints require an analyst session (Authorization: Bearer <JWT> from POST /v1/siu/auth/login) and are tenant-scoped to the analyst's carrier — one carrier's analysts can never read another carrier's data (ADR-037).

/v1/siu/auth/login                    Analyst login → JWT (8h)
/v1/siu/queue                         Case queue for SIU analysts
/v1/siu/cases/{id}                    Full case detail + scoring breakdown
/v1/siu/cases/{id}/decision           Accept / reject / pend decision
/v1/siu/cases/{id}/evidence-pack      Trigger PDF evidence pack generation
/v1/siu/rings · /rings/{id}           Ranked fraud-ring queue + ring detail
/v1/siu/typologies                    Learned fraud typologies
/v1/evidence-packs/{id}/download      Download court-defensible evidence PDF

---

Tech Stack

Layer	Technology
API	FastAPI (async), uvicorn
AI / Orchestration	Claude Sonnet 4.6 (scoring), LangGraph (clinical validation pipeline)
Observability	W&B Weave — every agent op traced
Database	PostgreSQL + asyncpg (prod), SQLite + aiosqlite (dev/test)
Background jobs	Celery + Redis
Encryption	AES-256-GCM field encryption (crypto.py)
Audit ledger	SHA-256-chained immutable ledger (ledger.py)
External registries	OIG LEIE (daily sync), NPPES NPI REST API (7-day cache)
Frontend	React 18 + TypeScript + Vite + Tailwind CSS (rybera-frontend/)
Migrations	Alembic (12 migrations: V2 → V13)
Containerisation	Docker + docker-compose (PostgreSQL, Redis, API, Worker, Flower)

---

HIPAA Compliance

Every enforcement point is documented in DECISIONS.md (ADR-001 through ADR-038):

• patient_id SHA-256 hashed at the first line of every API endpoint — raw value never stored
• clinical_note processed in memory only — stripped before any DB write
• Cache-Control: no-store on all responses (middleware-enforced)
• AES-256-GCM field encryption for PHI columns (FIELD_ENCRYPTION_KEY)
• Immutable audit ledger with hash chain (court-defensible, ADR-002)
• Per-carrier KMS-backed key derivation (carrier_kms.py, Phase 3)
• OIG LEIE exclusion check on every claim at ingest

---

Quickstart

Local development (SQLite, no Redis)

cp .env.example .env
# .env ships ready for local dev: ENVIRONMENT=development + SQLite.
#   - Set ANTHROPIC_API_KEY   → required for clinical validation / scoring
#   - Set FIELD_ENCRYPTION_KEY → required before ANY real PHI (dev runs without it)

pip install -r requirements.txt
python seed_db.py     # loads the demo carrier + sample claims — the dashboard is empty without it
python start.py       # serves the API and the prebuilt frontend on :8000

Open http://localhost:8000 — Rybera frontend loads automatically. Swagger UI at http://localhost:8000/docs.

The SIU portal fails closed (ADR-037): sign in with the seeded demo analyst — carrier carrier-archive-medicare, username demo.analyst, password RyberaDemo2026! (or your RYBERA_DEMO_ANALYST_PASSWORD). Create more analysts via POST /v1/siu/auth/register (open in development; bootstrap-token-gated in production).

Run the test suite (477 pass, 12 documented xfail — see KNOWN_ISSUES.md):

pytest -q

For live frontend development (hot-reload dev server on :5173, proxies to the API on :8000):

cd rybera-frontend && npm install && npm run dev

Production (Docker)

cp .env.example .env
# Set for production: ENVIRONMENT=production, a Postgres DATABASE_URL,
# FIELD_ENCRYPTION_KEY, SECRET_KEY, and CORS_ALLOWED_ORIGINS
# (comma-separated carrier portal URLs — no wildcard in production).

alembic upgrade head   # build/upgrade the Postgres schema to V13
docker compose up -d

Services: api (port 8000), worker (Celery), flower (port 5555), db (Postgres), redis.

---

API Reference

Fraud Scoring

POST /v1/claims/score
Header: X-Rybera-Carrier-Key: <carrier_key>

{
  "carrier_id": "carrier-abc",
  "external_claim_id": "CLM-2026-001",
  "provider_npi": "1234567890",
  "patient_id": "MBR-99999",          ← hashed at boundary
  "em_code_billed": "99215",
  "icd10_codes": ["M17.11"],
  "cpt_codes": ["99215", "20610"],
  "modifiers": [],
  "billed_amount": 485.00,
  "date_of_service": "2026-06-01",
  "clinical_note": "..."               ← never stored
}

→ {
    "composite_score": 82,
    "score_recommendation": "ESCALATE",
    "score_reliability": "reliable",
    "layer2_score": 71,
    "layer3_score": 88,
    "layer4_score": 65,
    ...
  }

Clinical Validation

POST /api/v1/validate-claim
Header: X-API-Key: <tenant_key>

→ {
    "validation_id": "...",
    "fraud_score": 74,
    "risk_level": "HIGH",
    "disposition_recommendation": "HOLD",
    ...
  }

---

Key Files

File	Purpose
server.py	Unified FastAPI entry point — all routes
database.py	SQLAlchemy async ORM — all schema models
scoring.py	Composite fraud score aggregator
ingest.py	Claim normalization + NPPES/LEIE enrichment
layer2_ai_inflation.py	AI scribe upcoding detection
layer3_provider_pattern.py	24-month provider behavior profiling
layer4_policyholder.py	Cross-provider policyholder patterns
layer5_graph.py	Network graph — collusion ring detection
validation_graph.py	Clinical validation pipeline (LangGraph)
fraud.py	Deterministic fraud-signal primitives (modifier / POS / volume)
phi_redaction.py	Clinical-note redaction for the audit trail (ADR-036)
siu_portal.py	SIU case management router
evidence_pack.py	ReportLab evidence PDF generation
ledger.py	SHA-256-chained immutable audit ledger
crypto.py	AES-256-GCM field encryption
carrier_kms.py	Per-carrier HKDF key derivation
external_registry.py	OIG LEIE + NPPES NPI registry
ai_scribe_detector.py	Clinical note AI-generation detection
clinical_agent.py	LLM-based MDM scoring agent
worker.py	Celery background tasks
start.py	Development launcher

---

Business Model

Rybera is priced on shared savings: 20% of overbilling recovered through Rybera's intervention. No upfront licensing — carriers pay only when Rybera catches fraud.

SHARED_SAVINGS_RATE=0.20 (configurable per carrier in .env).

---

Architecture Decisions

All non-obvious design choices are documented in DECISIONS.md (ADR-001 through ADR-038), covering database choice, immutable ledger design, HIPAA de-identification approach, encryption strategy, OIG sync method, provider profile window, real-time scoring latency constraints, NPPES caching, evidence-pack determinism, analyst-id-from-session auth (ADR-010), and the fraud-ring intelligence layer.