DeFiPulseAI

Multi-Agent DeFi Analytics Pipeline — 7 specialized agents, on-chain data ingestion across 9 chains, and MiMo-powered reasoning that turns raw transactions into analyst-grade reports.

---

What is this?

DeFiPulseAI is a production-grade multi-agent orchestration system that turns raw on-chain data into actionable DeFi intelligence. It uses MiMo (Xiaomi) as the primary reasoning engine and Claude as a cross-check backend, with the Hermes Agent framework orchestrating 7 specialized agents through configurable DAGs.

Five features, one pipeline

Feature	What it does	Example output
🐋 Smart Money Tracker	Identify alpha wallets, score their PnL, classify behavior	"0xabc… is a swing_trader (conf 0.87), alpha 0.74"
🚨 Liquidation Risk Monitor	Scan Aave/Compound positions, flag HF < 1.2	"0xdef…: HF 1.05 — URGENT, ETA to liq ~12 min"
🔍 Wallet Forensics	Deep multi-chain profile + LLM narrative	"Active 18 months across 4 chains; preferred protocol: Uniswap; risk: LOW"
🔓 Token Unlock Calendar	Upcoming unlocks + LLM-derived dump-risk scores	"ARB unlock in 5d: 2.3% supply, impact 78/100 — high risk"
💰 Yield Scanner	Cross-protocol APY compare + LP recommendations	"Pendle PT-USDe 14d: 22% APY, strong_yes — TVL stable, IL=0"

---

Architecture

graph LR
  classDef agent fill:#4f46e5,stroke:#fff,color:#fff;
  classDef llm fill:#dc2626,stroke:#fff,color:#fff;
  classDef io fill:#059669,stroke:#fff,color:#fff;

  A[IngestionAgent]:::agent --> B[TransformationAgent]:::agent
  B --> C[EnrichmentAgent]:::agent
  C --> D[AnalyticsAgent]:::agent
  D --> E[AlertAgent]:::agent
  D --> F[ReportingAgent]:::agent
  G[SchedulerAgent]:::agent -.-> A

  C -.uses.-> M[MiMo]:::llm
  C -.cross-check.-> CL[Claude]:::llm

  A -.fetches.-> AL[Alchemy / Helius / Etherscan]:::io
  A -.fetches.-> DL[DefiLlama / CoinGecko / TheGraph]:::io
  E -.sends.-> TG[Telegram / Discord / Slack]:::io
  F -.writes.-> RP[Markdown / JSON / PDF]:::io

Loading

The 7 agents

1. IngestionAgent (~450 LOC) — pulls data from Alchemy, Helius, Etherscan, DefiLlama, CoinGecko, The Graph
2. TransformationAgent (~500 LOC) — normalizes/decodes/dedupes; chain-aware tx classification
3. EnrichmentAgent (~400 LOC) — LLM-driven tagging, narrative generation, dump-risk scoring (uses MiMo + Claude)
4. AnalyticsAgent (~450 LOC) — Gini, HHI, impermanent loss, health factor, alpha scoring
5. AlertAgent (~370 LOC) — pattern matching → Telegram / Discord / Slack webhook fan-out
6. SchedulerAgent (~250 LOC) — recurring poll loops, DAG suggestions per feature
7. ReportingAgent (~280 LOC) — Markdown render + structured table + metric extraction

Total: 3,400+ LOC of agent code, 6 adapters for upstream data, 5 feature wrappers, and 170+ unit tests.

---

Quickstart

git clone https://github.com/<you>/DeFiPulseAI
cd DeFiPulseAI
pip install -e ".[dev]"
cp .env.example .env  # fill in keys

Python API

import asyncio
from defipulse import load_config
from defipulse.features import WalletForensics, LiquidationRiskMonitor

async def main():
    config = load_config()

    # 1. Deep forensic profile of a single wallet
    async with await WalletForensics.from_config(config) as wf:
        report = await wf.run(
            "0xd8da6bf26964af9d7eed9e03e53415d37aa96045",
            chains=["ethereum", "base"],
        )
        print(report.summary)

    # 2. Liquidation risk poll across a set of borrowers
    async with await LiquidationRiskMonitor.from_config(config) as lr:
        at_risk = await lr.at_risk(
            ["0xabc...", "0xdef..."], threshold=1.2,
        )
        for row in at_risk:
            print(f"⚠️ {row['wallet']}: HF {row['health_factor']:.2f}")

asyncio.run(main())

CLI

# wallet forensics
defipulse forensics 0xabc... --chains ethereum,base

# liquidation scan
defipulse liquidation 0xabc... 0xdef...

# top alpha wallets
defipulse smart-money --chains ethereum --top 10

# yield opportunities
defipulse yields --chains ethereum --min-tvl 5000000

# upcoming token unlocks
defipulse unlocks --days 30

---

Configuration

All settings flow through .env → defipulse.core.config.load_config():

# LLM providers
MIMO_API_KEY=...
MIMO_BASE_URL=https://9router.example.com/v1      # or SwiftRouter, or direct
MIMO_MODEL=mimo-v2-flash

ANTHROPIC_API_KEY=...                              # cross-check backend
CLAUDE_MODEL=claude-sonnet-4-6

# Data providers
ALCHEMY_API_KEY=...
HELIUS_API_KEY=...
ETHERSCAN_API_KEY=...
COINGECKO_API_KEY=...
THE_GRAPH_API_KEY=...

# Notifications (any subset)
TELEGRAM_BOT_TOKEN=...
TELEGRAM_CHAT_ID=...
DISCORD_WEBHOOK_URL=...
SLACK_WEBHOOK_URL=...

See .env.example for the full list.

---

Supported chains

Chain	EVM?	Native	Adapter
Ethereum	✅	ETH	Alchemy + Etherscan
Base	✅	ETH	Alchemy + Etherscan
Arbitrum	✅	ETH	Alchemy + Etherscan
Optimism	✅	ETH	Alchemy + Etherscan
Polygon	✅	POL	Alchemy + Etherscan
BSC	✅	BNB	Etherscan
Avalanche	✅	AVAX	Etherscan
Solana	❌	SOL	Helius
Hyperliquid	❌	USDC	(planned)

---

How the LLM is used

The EnrichmentAgent is where the LLM budget is spent. Each call has a narrow, structured contract:

Task	Prompt → Response	LLM
Wallet classification	"Given activity X, tag this wallet" → {tag, confidence, reason}	MiMo
Forensic narrative	Multi-chain summary → 2-paragraph profile	MiMo
Forensic cross-check	First-draft + raw evidence → "missed any flags?"	Claude
Unlock dump-risk score	event metadata → {impact_score: 0-100, rationale}	MiMo
Yield review	pool metadata → {recommendation, rationale}	MiMo

Why this matters: every LLM call has a tight JSON contract verified by complete_json(), so we never burn tokens on un-parseable output.

---

Testing

pytest tests/ -v          # all 170+ tests
pytest tests/ --cov       # with coverage
pytest -k "analytics"     # filter

Every external API is mocked via httpx.MockTransport — tests run offline and finish in under 5 seconds.

---

Project structure

DeFiPulseAI/
├── src/defipulse/
│   ├── __init__.py
│   ├── cli.py
│   ├── core/
│   │   ├── config.py
│   │   ├── types.py
│   │   ├── mimo_client.py
│   │   ├── claude_client.py
│   │   ├── hermes_runtime.py
│   │   └── pipeline.py
│   ├── adapters/
│   │   ├── base.py
│   │   ├── alchemy.py
│   │   ├── helius.py
│   │   ├── etherscan.py
│   │   ├── coingecko.py
│   │   ├── defillama.py
│   │   └── thegraph.py
│   ├── agents/
│   │   ├── ingestion.py
│   │   ├── transformation.py
│   │   ├── enrichment.py
│   │   ├── analytics.py
│   │   ├── alert.py
│   │   ├── scheduler.py
│   │   └── reporting.py
│   ├── features/
│   │   ├── smart_money.py
│   │   ├── liquidation_risk.py
│   │   ├── wallet_forensics.py
│   │   ├── token_unlock.py
│   │   └── yield_scanner.py
│   └── utils/
│       ├── cache.py
│       ├── logger.py
│       ├── retry.py
│       └── rate_limit.py
├── tests/                         # 170+ tests
├── examples/                      # runnable scripts
├── demo/                          # Next.js Vercel demo (web UI)
└── README.md

---

Roadmap

• 7-agent core pipeline
• 5 feature modules
• CLI + Python API
• 170+ unit tests
• Vercel demo site
• WebSocket support for real-time alerts
• More chains: Sei, Sui, TON, Aptos
• PDF report renderer
• Backtest engine for yield strategies
• On-chain identity resolution (ENS + Lens + Farcaster + SNS)
• Governance proposal summarizer (DAO module)

---

License

MIT — see LICENSE.

---

Acknowledgments

• Hermes Agent framework (in-tree implementation)
• Xiaomi MiMo
• Anthropic Claude
• DefiLlama, Alchemy, Helius, Etherscan, CoinGecko, The Graph