🏆 ENZIU — Universal Insurance Transparency Engine

AMD Developer Hackathon 2026 Submission
🤖 Track 1: AI Agents & Agentic Workflows
⚡ Powered by Llama 4 Scout 17B + Llama 3.3 70B on AMD Developer Cloud / Nscale

---

🎯 The Problem

Insurance policies are weapons of mass confusion. They span 50–100+ pages of intentional legal density. In 2024 alone, 8.8 million claimants found themselves on the wrong side of fine print they never understood.

You might think "ChatGPT or Claude can summarize PDFs" — and you'd be right. But summarization is not auditing. A summary tells you what the document says. An audit tells you what it means for you, scores it against objective criteria, flags hidden risks, and produces a reproducible grade that holds up to scrutiny.

That's ENZIU.

---

✨ The Solution

ENZIU is a multi-agent AI auditing system that doesn't just read your policy — it audits, scores, and grades it with the precision of an insurance auditor.

🔑 Key Differentiators

Feature	ChatGPT/Claude	ENZIU
Output	Summary	ENZIU Index (0–100) + Letter Grade
Consistency	Varies per request	Deterministic (±1 letter grade)
Method	General analysis	100+ point criteria across 3 dimensions
Privacy	Data may be stored	Zero server storage, client-side encryption
Citations	May hallucinate	Page-verified, excerpt-backed
Legal Standard	General knowledge	Dynamic; Based on document source

---

🏗️ Architecture: Multi-Agent System

ENZIU uses a two-phase agentic workflow where specialized AI agents coordinate to produce deterministic, reproducible results:

┌─────────────────────────────────────────────────────────────────┐
│                        PDF Upload                               │
│                    (io.BytesIO — Zero Disk)                     │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│  AGENT 1: ENZIU Extractor (Llama 4 Scout 17B — 890K context)    │
│  ─────────────────────────────────────────────────────────────  │
│  • Extracts 18 categories of structured facts (A–R)             │
│  • Performs legal risk scan as bad-faith attorney               │
│  • Records page citations for every finding                     │
│  • Output: Structured JSON facts (cached per session)           │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│  AGENT 2: ENZIU Auditor (Llama 3.3 70B — 131K context)          │
│  ─────────────────────────────────────────────────────────────  │
│  • Scores Clarity (0–30 pts): reading grade, jargon, navigation │
│  • Scores Coverage (0–40 pts): exclusions, waiting periods      │
│  • Scores Claims (0–30 pts): appeal rights, payout timeline     │
│  • Detects Red Flags: 2-source detection (finding + structural) │
│  • Generates 8 Insight Cards with page citations                │
│  • Output: Full ENZIU Report with Index & Grade                 │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│                    ENZIU Index Calculation                      │
│  ─────────────────────────────────────────────────────────────  │
│  base_score = clarity + coverage + claims (max 100)             │
│  enziu_index = base_score − red_flag_deductions (cap 40)        │
│  grade = A+ (90+) | A (80+) | B+ (75+) | B (70+) | ... | F (<50)│
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│              Client-Side Encrypted Storage (IndexedDB)          │
│  ─────────────────────────────────────────────────────────────  │
│  • AES-256-GCM encryption with PBKDF2 key derivation            │
│  • Recovery Vault keyed by SHA256(voucher_code)                 │
│  • Zero server-side persistence                                 │
└─────────────────────────────────────────────────────────────────┘

---

🔬 The ENZIU Index

The ENZIU Index is a proprietary 0–100 score that objectively evaluates insurance policies across three dimensions:

📐 Scoring Criteria (100+ Points)

1. Clarity Score (0–30 points)

• Reading Grade (8 pts): Flesch-Kincaid analysis across 3 samples
• Jargon Density (6 pts): Undefined legal terms count
• Definitions Completeness (6 pts): Defined vs capitalized terms ratio
• Passive Voice (5 pts): Passive construction ratio
• Navigability (5 pts): TOC, section numbering, cross-references

2. Coverage Score (0–40 points)

• Exclusion Volume (12 pts): Count and prominence of exclusions
• Waiting Period (8 pts): Presence, page location, days clarity
• Sub-Limit Transparency (8 pts): Location and clarity of sub-limits
• Pre-Existing Conditions (6 pts): Lookback period analysis
• Renewability & Cancellation (6 pts): Notice periods, grounds

3. Claim Efficiency Score (0–30 points)

• Filing Clarity (8 pts): Contact, forms, documentation, deadlines
• Appeal Rights (8 pts): Internal + external review availability
• Payout Timeline (7 pts): Explicit days commitment
• Dispute Resolution (7 pts): Regulator reference, arbitration flags

🚩 Red Flag Detection (Two-Source)

Source A — Finding-Triggered: Every risk finding from the Extractor that meets severity thresholds becomes a red flag with deduction values (1–10 points each).

Source B — Structural: Flags triggered by recorded facts independent of risk findings:

• no_internal_appeal (critical, 10 pts)
• sub_limits_buried (major, 6 pts)
• waiting_period_noncompliant (major, 5 pts)
• missing_sbc (minor, 4 pts)
• renewal_terms_absent (minor, 3 pts)
• no_regulator_reference (minor, 3 pts)

📊 Grade Bands

ENZIU Index	Grade	Preview
90–100	A+	High
80–89	A	High
75–79	B+	Medium
70–74	B	Medium
65–69	C+	Medium
60–64	C	Low
50–59	D	Low
<50	F	Low

Determinism Guarantee: Same policy → Same facts → Same grade (±1 letter grade band, e.g., B+ or B- for B).

---

🔒 Privacy-First Architecture

Zero Server Storage

# All PDF processing happens in memory
buffer = io.BytesIO(content)  # Never touches disk
doc = fitz.open(stream=buffer, filetype="pdf")
# Processing...
doc.close()  # Nothing persisted

Client-Side Encryption

Reports are encrypted in the browser using AES-256-GCM with keys derived via PBKDF2:

// web/lib/pdf-storage.ts
const salt = crypto.getRandomValues(new Uint8Array(16));
const iv = crypto.getRandomValues(new Uint8Array(12));
const key = await deriveSessionKey(sessionId, salt);
const ciphertext = await crypto.subtle.encrypt(
  { name: 'AES-GCM', iv: iv.buffer as ArrayBuffer },
  key,
  plaintext
);

Recovery Vault

Lost your voucher? Recover your report without email or PII:

// Store encrypted vault keyed by SHA256(voucher_code)
await storeRecoveryVault(voucherCode, { factSheet, extractedText, sessionId });

// Retrieve with just the voucher code
const data = await getRecoveryVault(voucherCode);

Remember that it is stored locally. So if you cleared your browser or used a new machine, it'll be unrecoverable.

---

⚡ Llama Stack Integration

ENZIU leverages the full Llama family for optimal performance:

Agent	Model	Context	Purpose
Extractor	Llama 4 Scout 17B	890K	Fact extraction, legal risk scan
Auditor	Llama 3.3 70B	131K	Scoring, grading, insight generation

Why Two Models?

• Scout 17B: Massive 890K context handles entire policies (100+ pages) in one pass
• Llama 3.3 70B: Superior reasoning for complex scoring and legal analysis
• Temperature 0.0: Deterministic output for reproducible grades

Inference Infrastructure

# api/app/config.py
inference_api_base: str = "https://inference.api.nscale.com/v1"
inference_model: str = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
auditor_model: str = "meta-llama/Llama-3.3-70B-Instruct"

Powered by AMD Developer Cloud for development and NScale for production scaling.

---

🚀 Getting Started

Prerequisites

• Python 3.11+
• Node.js 18+
• AMD Developer Cloud account (for GPU access)
• NScale API key (for inference)

Backend Setup

# Clone repository
git clone https://github.com/shannja/enziu.git
cd enziu/api

# Install dependencies
poetry install

# Configure environment
cp ../.env.example ../.env
# Edit .env with your API keys

# Run development server
poetry run uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

Frontend Setup

cd web

# Install dependencies
npm install

# Run development server
npm run dev

Environment Variables

# Inference (NScale)
INFERENCE_API_KEY=your_nscale_api_key
INFERENCE_API_BASE=https://inference.api.nscale.com/v1
INFERENCE_MODEL=meta-llama/Llama-4-Scout-17B-16E-Instruct
AUDITOR_MODEL=meta-llama/Llama-3.3-70B-Instruct

# Paddle Billing
PADDLE_ENV=sandbox
PADDLE_CLIENT_TOKEN=your_client_token
PADDLE_WEBHOOK_SECRET=whsec_your_secret
PADDLE_PRODUCT_ID=pro_xxxxx

# Voucher System
VOUCHER_HMAC_SECRET=your_hmac_secret

# Security
API_SECRET_KEY=openssl rand -hex 32

---

📊 Live Demo

Try ENZIU live on Vercel:

👉 enziu.vercel.app

---

🛠️ Tech Stack

Backend

• FastAPI — Async Python web framework
• PyMuPDF — In-memory PDF extraction
• HTTPX — Async LLM API client
• Bcrypt — Voucher passphrase hashing
• SlowAPI — Rate limiting

Frontend

• Next.js 15 — React 19 with App Router
• TypeScript — Type-safe development
• Tailwind CSS — Utility-first styling
• Radix UI — Accessible components
• Framer Motion — Animations
• IndexedDB — Client-side encrypted storage

AI/ML

• Llama 4 Scout 17B — Fact extraction (890K context)
• Llama 3.3 70B — Policy auditing (131K context)
• NScale Inference API — Production serving
• AMD Developer Cloud — Development & benchmarking

Infrastructure

• Vercel — Frontend hosting
• AMD Developer Cloud / Nscale — GPU compute
• Paddle — Payment processing

---

📝 API Reference

Extract & Audit Policy

POST /api/extract
Content-Type: multipart/form-data

file: <pdf_file>

Response:

{
  "session_id": "uuid",
  "extracted_text": "[{\"page_number\": 1, \"text\": \"...\"}]",
  "grade": {
    "overall": "B+",
    "clarity": "A",
    "coverage": "B",
    "claimsEfficiency": "C+"
  },
  "topRisk": "No internal appeal process",
  "redFlags": ["No internal appeal process"],
  "summary": "This policy covers...",
  "score_preview": "medium",
  "policy_type": "health",
  "carrier_name": "Example Insurance",
  "full_report": { /* Complete ENZIU report */ }
}

Validate Voucher

POST /api/voucher/validate
Content-Type: application/json

{
  "code": "ENZ-ABCD-EFGH-IJKL-MN",
  "passphrase": "my-secure-passphrase"
}

Policy Audit (Cache Lookup)

POST /api/policy/audit
Content-Type: application/json

{
  "session_id": "uuid",
  "extracted_text": "[...]"
}

---

🤝 Contributing

ENZIU is open source under the MIT License. Contributions welcome!

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing)
5. Open a Pull Request

---

📄 License

MIT License — see LICENSE for details.

---

🙏 Acknowledgments

• AMD Developer Cloud — GPU compute credits and support
• Meta AI — Llama 4 and Llama 3.3 open-source models
• NScale — Production inference infrastructure
• Hugging Face — Model hub and deployment platform
• lablab.ai — Hackathon platform

---

📬 Contact

• GitHub: github.com/shannja/enziu
• Website: enziu.vercel.app

---

Built with ⚡ by eseyem Team for AMD Developer Hackathon 2026

🤖 Track 1: AI Agents & Agentic Workflows

This is a hackathon project only for Lablab.AI's AMD Developer Hackathon. Not a real operating business.