FinAgentX

💡 Inspiration

Traditional lending is broken—overcollateralized DeFi excludes most users, while traditional banks rely on opaque, slow, and biased credit systems. We wanted to explore a world where AI agents themselves become financial institutions—autonomous, transparent, and always online.

FinAgentX was inspired by the idea of a self-sovereign economic actor that can evaluate trust, take risk, and manage capital without human intervention.

⚙️ What it does

FinAgentX is an autonomous on-chain AI bank that:

Evaluates borrower risk using an ML ensemble model

Issues undercollateralized USDT loans based on creditworthiness

Dynamically adjusts interest rates using risk + uncertainty

Uses an LLM to negotiate loan terms and generate explanations

Monitors loans and automatically collects repayments

Continuously learns from on-chain outcomes to improve decisions

It replaces traditional trust with data, models, and verifiable on-chain execution.

🏗️ How we built it

We built FinAgentX as a full-stack autonomous system:

Smart Contracts (Solidity + Hardhat):

Loan lifecycle, lending pool, and DID-based credit identity

Autonomous Agent (Node.js):

12-step reasoning loop handling evaluation, approval, and monitoring

Integrated with Tether WDK for self-custodial transaction execution

ML Engine (Python + Scikit-learn):

Ensemble of Linear, Ridge, Lasso, and Random Forest models

Real-time inference via FastAPI

Continuous learning from on-chain repayment data

LLM Layer (Ollama + Gemma 2B):

Negotiates loan terms and explains decisions

Frontend (React):

Dashboard, loan requests, risk visualization, and agent monitoring

Everything runs together as a fully autonomous financial loop.

⚠️ Challenges we ran into

Designing undercollateralized lending without introducing high default risk

Balancing ML accuracy vs. on-chain latency

Integrating LLM reasoning into deterministic financial decisions

Managing agent autonomy + safety (avoiding malicious or irrational behavior)

Creating a realistic synthetic dataset for initial training

Handling cross-stack communication between Node.js, Python, and Solidity

🏆 Accomplishments that we're proud of

Built a fully autonomous 12-step AI agent loop

Achieved real on-chain execution using Sepolia USDT (no mocks)

Designed a hybrid ML + LLM decision system

Implemented continuous learning from blockchain data

Created a working prototype of trustless credit scoring

Enabled agent-to-agent financial interactions

📚 What we learned

Autonomous agents can act as economic entities, not just tools

Combining ML (quantitative) and LLMs (qualitative reasoning) is powerful

On-chain systems demand determinism, safety, and transparency

Continuous learning is essential for adaptive financial systems

Building across Web3 + AI + backend + frontend requires tight integration

🔮 What’s next for FinAgentX

Deploy on mainnet / Layer 2 for real-world usage

Replace synthetic data with real borrower histories

Introduce reputation-based credit networks (on-chain DID graph)

Add multi-agent coordination (AI banks interacting with each other)

Improve LLM reasoning with larger or fine-tuned models

Expand to multi-asset lending beyond USDT

Build a fully decentralized governance layer

Built With

• sepolia
• solidity