AI-Yield-Architect

Demo

About

The problem it solves AI Yield Architect tackles four major problems for DeFi users: * **Overwhelming Complexity:** The DeFi yield landscape is fragmented and complex. Users must manually track APYs, risks, and strategies across dozens of protocols like Aave, Morpho, Spark, and Yearn. **AI Yield Architect simplifies this by aggregating them into a single, "deposit-and-forget" ERC-4626 vault.** * **Static, Time-Consuming Management:** Manually rebalancing positions to chase the best yields is time-consuming, gas-intensive, and impractical for most users. **Our platform makes this task easier by using an AI Strategy Engine (powered by Google Gemini) to continuously analyze market data 24/7 and provide optimal, real-time allocation recommendations.** * **Opaque Risk:** High APYs often hide significant risk (e.g., smart contract vulnerabilities, low liquidity). **AI Yield Architect makes yield farming safer by providing AI-driven risk assessments, analyzing protocol TVL, historical performance, and other metrics to recommend risk-adjusted strategies, not just the highest-APY "degen" plays.** * **Friction Between Profit and Purpose:** Most yield farming is purely extractive. Users profit from the ecosystem without a simple way to support the public goods (core devs, infrastructure) that make it possible. **We solve this by integrating Octant v2, automatically and transparently donating 4.2% of all earned yield to public goods, allowing users to earn competitive returns while simultaneously funding the future of Ethereum.**

How it's made

Challenges we ran into Building a multi-protocol, AI-driven vault that also donates yield is a complex task. The three biggest hurdles we faced were: 1. **The ERC-4626 `totalAssets()` Calculation:** Our entire `AIYieldVault` is built on the ERC-4626 standard, where the vault's share price is derived from the `totalAssets()` function. Our challenge was that our vault doesn't hold the assets directly; they are all deposited in external protocols (Aave, Morpho, Spark, Kalani). * **The Bug:** A simple `balanceOf(USDC)` check on our vault returned 0, as all the funds were in the adapters. This meant the vault's share price never increased, and depositors couldn't redeem their accrued yield. * **The Solution:** We had to implement a standardized `getBalance()` function in each protocol adapter (`AaveAdapter.sol`, `MorphoAdapter.sol`, etc.). The main `totalAssets()` function now loops through all active adapters, calls `getBalance()` on each, and sums the results. This was especially complex for the `YearnAdapter`, as we had to trust *its* internal calculations to get an accurate value for our vault-within-a-vault. 2. **Implementing the Octant v2 "Yield Skimming" Strategy:** Our 4.2% public goods donation is not a simple fee. It's a *yield-donating strategy* as defined by Octant v2. * **The Hurdle:** We couldn't just take 4.2% of a deposit. We had to design our `harvest()` function to *only* skim the *profit* (yield) generated across all four protocols, without ever touching the user's principal. * **The Solution:** This required a complex `harvest()` function that, when triggered, calculates the vault's total assets *minus* its total liabilities (original deposits). This resulting "profit" is then split, with 4.2% being routed to the Octant v2 donation address and the rest being compounded back into the vault for depositors. Getting this "profit-only" calculation right was a significant smart contract security challenge. 3. **Connecting Real-Time AI to On-Chain Data:** The AI Strategy Engine is not a gimmick. It needs to provide real, actionable advice based on live market conditions, which presented a major infrastructure challenge. * **The Hurdle:** You cannot call an LLM (like Google Gemini) directly from a smart contract. Furthermore, LLMs don't understand raw on-chain data (like APYs in 27-decimal `ray` format). * **The Solution:** We had to build a serverless backend using edge functions (`live-protocol-data`) to act as a crucial middle layer. This backend constantly: 1. Pulls raw data from all 4 protocols (Aave, Morpho, etc.). 2. Cleans and pre-processes this data into a simple, human-readable format (e.g., "Aave APY: 8.45%"). 3. Feeds this clean data to the Gemini AI as part of its prompt context. 4. Receives the AI's natural language recommendation and streams it to the frontend. This data pipeline was the most difficult *off-chain* component to build, as stale or badly formatted data would cause the AI to give useless or dangerous advice.