DeFiGuard - Multi-Agent DeFi Protocol Intelligence

Real-time DeFi protocol risk analysis powered by multi-specialist LLM agents on AMD ROCm

---

Architecture

                          DeFiGuard Multi-Agent Pipeline
  ┌──────────┐   ┌──────────┐   ┌─────────────────────────────────┐   ┌───────────┐   ┌────────┐
  │  Input    │──▶│ Chunking │──▶│         Agent Fan-out           │──▶│ Synthesis │──▶│ Report │
  │  Sources  │   │ Engine   │   │                                 │   │  Agent    │   │  PDF   │
  └──────────┘   └──────────┘   │  ┌─────────┐  ┌─────────────┐  │   └───────────┘   └────────┘
  - Token docs   - Smart       │  │  Risk    │  │ Tokenomics  │  │   - Consensus
  - Gov forums     contracts   │  │  Agent   │  │   Agent     │  │     scoring
  - On-chain     - Forum      │  └─────────┘  └─────────────┘  │   - Conflict
    data           posts      │  ┌─────────┐  ┌─────────────┐  │     resolution
  - Audit        - News       │  │Governance│  │  Thesis     │  │   - Final
    reports        articles   │  │  Agent   │  │   Agent     │  │     narrative
                 - Proposals  │  └─────────┘  └─────────────┘  │
                  ──────────  └─────────────────────────────────┘
                              4 specialist agents run in parallel
                              Each agent sees full chunked context
                              consuming 20-80K tokens per analysis

Why AMD ROCm?

DeFiGuard's multi-agent architecture is a natural fit for AMD ROCm GPUs, and here's why:

Memory Requirements

Each of our 4 specialist agents processes the full chunked context of a DeFi protocol simultaneously. A single analysis run loads:

• Smart contract source code (5-30K tokens)
• Governance forum history (10-50K tokens)
• On-chain metrics and audit reports (5-20K tokens)

The MI300X's 192GB HBM3 allows us to keep multiple agent contexts in VRAM without paging, enabling true parallel inference across all 4 agents.

Open Ecosystem

• No vendor lock-in: ROCm's open-source stack means we can profile, debug, and optimize every layer
• vLLM/TGI native support: Drop-in compatibility with existing LLM serving infrastructure
• PyTorch-native: Our entire agent orchestration layer runs on standard PyTorch with ROCm backend

Cost Efficiency

At scale, ROCm GPUs deliver 2-3x better token/$ compared to equivalent NVIDIA hardware, which is critical for our use case where a single research desk can burn 6M+ tokens/day.

---

Token Burn Estimates

DeFiGuard's multi-agent architecture naturally consumes large volumes of tokens. Each analysis fans out to 4 agents, each processing the full chunked context independently.

Scenario	Items	Chunks	Agents	Tokens/Day
Single deep dive	1	4	4	~100K
Daily analyst	5	20	4	~500K
Research desk	15	60	4	~2M
Continuous monitoring	25+	100+	4+chat	~6M+

Why so many tokens? Each agent receives the full context window. A 5-chunk analysis with 4 agents = 20 LLM calls minimum. Add synthesis, follow-ups, and monitoring, and a research desk easily hits 2M tokens/day.

---

Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Configure environment
cp .env.example .env
# Edit .env with your API keys and ROCm settings

# 3. Start the server
uvicorn app.main:app --host 0.0.0.0 --port 8000

# 4. Open the dashboard
# Navigate to http://localhost:8000 in your browser

Environment Variables

# LLM Backend (ROCm)
LLM_BASE_URL=http://localhost:8080/v1
LLM_MODEL=defiguard-70b
LLM_API_KEY=your-key-here

# Data Sources
DEFILLAMA_API_KEY=...
ETHERSCAN_API_KEY=...
SNAPSHOT_SPACE_ID=...

# Application
LOG_LEVEL=INFO
MAX_CONCURRENT_AGENTS=4
CHUNK_SIZE=4096

---

API Endpoints

Analysis

Method	Endpoint	Description
POST	/api/analyze	Trigger full multi-agent analysis of a protocol
GET	/api/analysis/{id}	Retrieve analysis results and report
GET	/api/analysis/{id}/status	Check analysis progress

Protocols

Method	Endpoint	Description
GET	/api/protocols	List all tracked protocols
GET	/api/protocols/{slug}	Get protocol details and latest risk score
GET	/api/protocols/{slug}/history	Historical risk scores

Reports

Method	Endpoint	Description
GET	/api/reports	List generated reports
GET	/api/reports/{id}	Download report (PDF/JSON)
GET	/api/reports/{id}/agents	Per-agent breakdown of report

Monitoring

Method	Endpoint	Description
POST	/api/monitor	Add protocol to continuous monitoring
DELETE	/api/monitor/{slug}	Remove from monitoring
GET	/api/monitor/alerts	Get active alerts

---

Agent Descriptions

Risk Agent

Analyzes smart contract security, audit history, TVL trends, and exploit vectors. Cross-references known vulnerability databases and flags patterns similar to past exploits.

Tokenomics Agent

Models supply dynamics, emission schedules, vesting cliffs, and economic incentive alignment. Identifies potential death spirals, inflation risks, and misaligned token mechanics.

Governance Agent

Scans Snapshot proposals, on-chain voting records, and forum discussions. Detects governance attacks, whale concentration, quorum risks, and proposal manipulation.

Thesis Agent

Synthesizes qualitative narrative: team credibility, competitive positioning, ecosystem partnerships, and market sentiment. Provides the "story" that contextualizes the quantitative signals.

Synthesis Agent

Not a domain specialist — this agent reconciles outputs from all 4 agents, resolves conflicting signals, and produces the final consensus risk score and executive summary.

---

Screenshots

Dashboard	Analysis View	Report PDF

---

Contributing

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

Please read CONTRIBUTING.md for detailed guidelines on code style, testing, and agent development.

---

License

This project is licensed under the MIT License — see the LICENSE file for details.

Copyright (c) 2026 Owonftt