🏗️ Dynamic RWA Yield Router
Mantle Turing Test Hackathon 2026 — Phase 2 "AI Awakening" — AI × RWA Track
An autonomous AI agent that dynamically allocates capital across Mantle's Real World Asset (RWA) stack using reinforcement learning, with all decisions recorded on-chain via ERC-8004 agent identity NFTs.
📊 What It Does
The Dynamic RWA Yield Router manages a portfolio across three core RWA assets on Mantle L2:
| Asset | Type | Target APY | Risk Profile |
|---|---|---|---|
| USDY (Ondo Finance) | Tokenized US T-Bills | ~4.25% | Low — USD-pegged, regulated |
| mETH (Mantle LSP) | Liquid Staked ETH | ~3-5% | Medium — ETH price exposure |
| MI4 (Mantle Index Four) | Tokenized Index Fund | ~5-8% | Medium — diversified basket |
The agent uses a PPO (Proximal Policy Optimization) reinforcement learning policy to learn optimal allocation weights given real-time market conditions, yield rates, and risk signals.
🏛️ Architecture
┌──────────────────────────────────────────────────────────────────┐
│ DYNAMIC RWA YIELD ROUTER │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ OBSERVE │→│ REASON │→│ PLAN │→│ AUTHORIZE│ │
│ │ │ │ │ │ │ │ │ │
│ │DeFiLlama│ │RL Policy │ │Rebalance │ │ Risk │ │
│ │CoinGecko│ │ (PPO) │ │ Trades │ │ Manager │ │
│ │Mantle │ │ │ │ │ │ │ │
│ │RPC/FRED │ │ │ │ │ │ │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
│ ↓ ↓ │
│ ┌──────────┐ ┌──────────┐ │
│ │ EXECUTE │←───────────────────────────│ VERIFY │ │
│ │ │ │ │ │
│ │Unsigned │ Mantle L2 (Chain 5000) │On-chain │ │
│ │Tx Builder│ Fluxion / Agni DEX │Confirm │ │
│ └──────────┘ └──────────┘ │
│ │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ ON-CHAIN CONTRACTS │ │
│ │ YieldRouterAgent.sol │ AgentIdentity8004.sol │ │
│ │ (Allocation Records) │ (ERC-8004 Reputation) │ │
│ │ RiskRegistry.sol │ (Risk Parameters) │ │
│ └─────────────────────────────────────────────────────────┘ │
└──────────────────────────────────────────────────────────────────┘
6-Stage Agent Pipeline
- OBSERVE — Fetch real-time yield rates, prices, and macro data from DeFiLlama, CoinGecko/Bybit, Mantle RPC, and FRED
- REASON — RL policy (PPO with actor-critic MLP) predicts optimal allocation weights
- PLAN — Compute required trades if weights drift beyond threshold (5%)
- AUTHORIZE — Multi-layer risk manager validates: depeg detection, volatility guard, position limits, drawdown protection, circuit breaker
- EXECUTE — Construct unsigned transaction payloads (construct-only safety: never holds private keys)
- VERIFY — Confirm execution, update portfolio state, record on-chain via ERC-8004
🧠 RL Model
Algorithm: Proximal Policy Optimization (PPO)
State Space (18-dimensional):
- Yield rates: USDY APY, mETH APY, MI4 APY, Aave supply APYs
- Prices: ETH, BTC, MNT (normalized)
- Macro: Fed funds rate, BTC dominance
- Risk: USDY peg, mETH peg, ETH 30d volatility
- Portfolio: current weights [USDY, mETH, MI4]
Action Space (3-dimensional, continuous):
- Raw logits → softmax → portfolio weights [USDY, mETH, MI4]
- Position limits enforced: 5-60% per asset
Reward Function:
reward = sharpe_ratio × 100
- drawdown × 50
- depeg_penalty
- rebalance_cost × 1000
Training Environment: Custom Gymnasium env with synthetic yield processes (Ornstein-Uhlenbeck for mean-reverting yields, GBM for prices).
🛡️ Risk Management
| Layer | Description | Trigger |
|---|---|---|
| Depeg Detection | Monitors USDY/USD and mETH/ETH pegs | USDY > 0.5%, mETH > 2% deviation |
| Volatility Guard | Reduces risky assets in high-vol regimes | ETH 30d vol > 80% |
| Concentration Limits | Enforces min/max position sizes | Any asset > 60% or < 5% |
| Smart Contract Risk | Weights by protocol audit status | Per-protocol risk scores |
| Drawdown Protection | Circuit breaker on losses | Portfolio DD > 10% |
| Emergency Exit | Flee to USDC on critical risk | Severe depeg or compounding risks |
📜 Smart Contracts
YieldRouterAgent.sol
- Records allocation decisions on-chain (audit trail)
- Enforces position limits and rebalance intervals
- Circuit breaker mechanism (OPEN/CLOSED/HALF_OPEN)
- Operator authorization for ERC-4337 wallets
AgentIdentity8004.sol (ERC-8004)
- Soulbound agent identity NFT (non-transferable)
- Capability registration and verification
- Attestation system (performance audits, security reviews)
- On-chain reputation scoring (0-10000 bps)
- Decision history references with outcome scoring
RiskRegistry.sol
- Per-asset risk parameters (max/min weights, depeg thresholds)
- Global risk parameters (max drawdown, slippage, gas limits)
- Depeg event tracking and resolution
- Allocation validation function
🚀 Quick Start
# Clone
git clone https://huggingface.co/muthuk1/mantle-rwa-yield-router
cd mantle-rwa-yield-router
# Install dependencies
pip install -r requirements.txt
# Run demo (3 cycles, live data)
python scripts/demo.py
# Train RL agent (optional, uses synthetic data)
python scripts/train.py
# Run continuous agent
python -m agent.main --mode run --interval 3600 --capital 100000
# Run tests
pytest tests/ -v
Environment Variables
cp .env.example .env
# Edit .env with your configuration
📂 Project Structure
mantle-rwa-yield-router/
├── agent/
│ ├── main.py # 6-stage orchestrator pipeline
│ ├── data_pipeline.py # Real-time yield & price data aggregation
│ ├── rl_optimizer.py # PPO agent + Gymnasium environment
│ ├── risk_manager.py # Multi-layer risk management
│ ├── executor.py # Unsigned transaction builder
│ └── strategy_reporter.py # LLM strategy letter generator
├── contracts/
│ ├── YieldRouterAgent.sol # On-chain allocation router
│ ├── AgentIdentity8004.sol # ERC-8004 agent identity NFT
│ └── RiskRegistry.sol # On-chain risk parameters
├── telegram_bot/
│ └── bot.py # Telegram UI (/status, /yields, /risk)
├── config/
│ └── constants.py # Contract addresses, ABIs, parameters
├── scripts/
│ ├── demo.py # Demo runner
│ └── train.py # RL training script
├── tests/
│ └── test_agent.py # 25 unit + integration tests
├── requirements.txt
├── .env.example
└── README.md
🔑 Key Design Decisions
Construct-Only Safety — The agent constructs unsigned transaction payloads but NEVER holds private keys. Execution requires external signing via ERC-4337 wallet or multisig.
Python-Native — Built in Python (not TypeScript) because RL/ML libraries (PyTorch, stable-baselines3, Gymnasium) are Python-native.
Fluxion DEX Primary — Fluxion is Mantle's native RWA-focused AMM, primary routing for RWA token swaps.
Risk-Parity Adaptive — Portfolio construction adapts to risk regimes. High volatility → more USDY (safe haven). Depeg → emergency exit to USDC.
On-Chain Audit Trail — Every allocation decision is recorded on-chain with content hashes linking to IPFS-stored strategy reports.
ERC-8004 + ERC-4337 — Agent identity via ERC-8004 (reputation, capabilities, attestations) + account abstraction via ERC-4337 (spending limits, session keys).
📊 Contract Addresses (Mantle Mainnet)
| Contract | Address |
|---|---|
| USDY (Ondo) | 0x5bE26527e817998A7206475496fDE1E68957c5A6 |
| mETH | 0xcDA86A272531e8640cD7F1a92c01839911B90bb0 |
| USDC | 0x09Bc4E0D864854c6aFB6eB9A9cdF58aC190D0dF9 |
| WMNT | 0x78c1b0C915c4FAA5FffA6CAbf0219DA63d7f4cb8 |
| Fluxion Router | 0x5628a59df0ecac3f3171f877a94beb26ba6dfaa0 |
| Agni Router | 0x319B69888b0d11cEC22caA5034e25FfFBDc88421 |
| Aave V3 Pool | 0x458F293454fE0d67EC0655f3672301301DD51422 |
🏆 Hackathon Track
Mantle Turing Test Hackathon 2026 — Phase 2 "AI Awakening"
- Track: AI × RWA
- Focus: Autonomous AI agents managing tokenized real-world assets
- Key Technologies: RL optimization, ERC-8004 agent identity, Mantle L2
📝 License
MIT
Built with 🤖 for the Mantle Turing Test Hackathon 2026