-
-
Logo
Inspiration DeFi yield farming is complex, time-consuming, and requires constant monitoring. Users miss opportunities, make suboptimal decisions, and struggle with gas costs. We wanted to build an AI agent that could autonomously manage USDT positions on Aave V3, making institutional-grade decisions 24/7 without human intervention. By combining Tether's WDK SDK with AI decision-making, we created a truly self-custodial, autonomous yield optimizer.
What it does YieldKernel is an autonomous AI agent that discovers real-time yields by fetching live USDT supply APY from Aave V3 on-chain data. It assesses risk using multi-factor scoring that evaluates protocol safety, liquidity, and APY thresholds. The system optimizes gas by calculating costs versus expected returns to ensure profitability.
For execution, it uses Tether's WDK SDK for self-custodial transaction handling. The agent monitors autonomously with 24-hour recheck cycles running in the background, independent of user connections. It leverages Groq AI to analyze market conditions and recommend optimal strategies.
The agent operates completely autonomously - users simply connect their wallet, specify an amount, and the AI handles everything else.
How we built it We built the frontend with vanilla JavaScript and Firebase Hosting. The backend runs on Node.js with Express and Firebase Functions. For blockchain interactions, we use Tether's WDK SDK, specifically the wallet and Aave lending protocol modules. The AI layer uses Groq's API with the llama-3.3-70b-versatile model. Data comes from Aave V3 on-chain queries, Chainlink price feeds, and public Ethereum RPCs.
The architecture has several key components. The self-custodial wallet layer uses WDK to manage HD wallets with BIP-39 mnemonics. Real-time data comes from direct on-chain queries to Aave V3 contracts for USDT reserve data. The risk engine implements multi-factor scoring covering protocol risk, liquidity, APY, and gas economics. The strategy engine supports balanced, aggressive, and conservative strategies with rebalancing logic. An autonomous monitor runs as a background service with 24-hour cycles. The AI decision layer uses Groq to analyze yields, risk, and gas costs.
We use WDK for all transaction execution including supply, withdraw, and approve operations. Ethers.js handles only read-only operations like Chainlink price feeds and RPC status checks. There's no mock data - everything comes from real on-chain sources. We implemented graceful error handling with proper fallbacks and deployed it for production use with real users.
Challenges we ran into The WDK SDK is in beta, so documentation was sparse. We had to reverse-engineer the API by reading source code and experimenting with different approaches.
Public RPC endpoints kept hitting rate limits or requiring authentication. We solved this by implementing a multi-endpoint fallback system with ethereum.publicnode.com, eth.llamarpc.com, and others.
We initially used "yield" as a variable name, which is a JavaScript reserved keyword. This caused deployment failures on Render. We fixed it by renaming to "yieldData".
Getting accurate gas estimates was tricky. We started with placeholder values but switched to WDK's quoteSupply method for real gas quotes.
Implementing true background monitoring that runs independently of user connections required careful state management and error handling to prevent memory leaks.
Fetching all Aave V3 reserves (63 assets) caused timeouts. We optimized by fetching only USDT data, reducing RPC calls by 98%.
Accomplishments that we're proud of We're running on 100% real data - no mock data anywhere. All yields, prices, and gas estimates come from on-chain sources. The background monitoring runs every 24 hours regardless of user activity, providing true autonomy. Users maintain full control of their funds through WDK wallets.
The system is production ready and deployed at https://yieldkernel-app.web.app with real users. We focused exclusively on USDT to meet the hackathon requirements. Groq AI makes intelligent decisions based on real-time market data. The system only rebalances when gas costs are justified by yield improvements. The 24-hour recheck cycles minimize unnecessary transactions and gas exposure, making it institutional grade.
What we learned We gained deep understanding of Tether's WDK SDK for self-custodial DeFi. We learned how to efficiently query Aave V3 contracts for real-time reserve data. Combining LLMs with DeFi requires careful prompt engineering and function calling. Production deployment taught us the importance of error handling, logging, and graceful degradation. We learned how to calculate break-even points for rebalancing decisions. Building truly autonomous agents that operate 24/7 without supervision was a valuable experience.
What's next for YieldKernel In the short term, we're adding multi-chain support for USDT on Arbitrum, Optimism, and Base. We'll build Telegram and Discord bots for notifications. Historical performance tracking and analytics are coming next, along with advanced strategies like leverage and delta-neutral positions.
Medium term plans include support for other Tether tokens on different chains. We'll integrate with more lending protocols like Compound V3 and Spark. Social features like copy trading and leaderboards are on the roadmap. A mobile app with push notifications is planned.
Long term, we're building DAO governance for strategy parameters. An institutional API will provide programmatic access. We'll add cross-protocol yield aggregation and automated tax reporting and compliance tools.
Our vision is to become the go-to autonomous yield optimizer for Tether holders, managing billions in USDT across multiple chains and protocols with institutional-grade safety and efficiency.
Log in or sign up for Devpost to join the conversation.