Alpha Street

The hedge fund industry manages $4.5 trillion, but 87% of people can't access it - accreditation barriers, $1M+ minimums, and multi-year lock-ups keep retail investors out. DeFi vaults solved access but introduced a fatal problem: strategy transparency. Every on-chain vault's strategy is public bytecode - visible, forkable, front-runnable. The moment a profitable strategy goes on-chain, it gets cloned and the alpha decays.

Alpha Street solves this with Chainlink CRE Confidential Compute. Investors deposit into an ERC-4626 vault - no minimums, no accreditation, no lock-ups. A CRE HTTP-triggered workflow executes each trading cycle: it reads vault state via EVM Read, fetches live market data via Confidential HTTP, runs AI inference for trading decisions via Confidential HTTP, and records trades on-chain. The critical innovation: both market data retrieval and AI reasoning execute inside the TEE enclave. The algorithm prompt, the market context, and Claude's full reasoning chain never leave the enclave - not to node operators, not to investors, not to anyone. Only the final trade action (buy/sell/hold, token pair, amount, confidence) exits.

Investors get full performance transparency - NAV history, trade log, share price - like a Bloomberg terminal. But they cannot deduce the strategy. This makes front-running impossible (you can't front-run an algorithm you can't see), prevents strategy cloning, and eliminates alpha decay. CRE is the sole executor - an onlyForwarder modifier ensures no human can directly trade fund assets. The protocol achieves what no existing system can: open access + verifiable performance + hidden strategy.

The contract layer is intentionally lean - just two Solidity contracts deployed on Avalanche Fuji with Foundry. AlphaVault is an ERC-4626 tokenized vault that handles investor deposits and withdrawals, tracks NAV as reported by CRE, and stores a history of every trade the AI has executed. TestUSDC is a simple mintable stablecoin for the hackathon demo. The key design decision was making AlphaVault also serve as the CRE report receiver - when CRE sends a report, the vault's onReport function decodes it and records both the trade and the updated NAV in one transaction.

The CRE workflow is the heart of the project. It's a single HTTP-triggered workflow written in TypeScript. Each invocation runs four steps in sequence: (1) EVM Read - reads totalAssets, currentNAV, and tradeCount from AlphaVault as a closed feedback loop so the AI always sees the fund's real current state; (2) Confidential HTTP - fetches live ETH/BTC/AVAX prices from CoinGecko inside the enclave; (3) Confidential HTTP - sends the market data and vault state to Claude (API key sealed as a Vault DON secret) and receives a structured JSON response: action (BUY/SELL/HOLD), token pair, amount, confidence score, and updated NAV; (4) EVM Write - ABI-encodes the decision and writes it to AlphaVault.onReport() via the Keystone forwarder. If the decision is HOLD, the workflow exits early without writing anything on-chain - no wasted gas. Because the AI model runs at temperature 0 with a fixed JSON schema, all DON nodes independently reach byte-identical decisions from identical inputs, so no explicit consensus voting is required.

We overrode the ERC-4626 totalAssets() function to return the CRE-reported NAV instead of the raw token balance. This way, share prices automatically reflect the fund's true value as computed by CRE, and investors can deposit or withdraw at accurate prices.

The frontend is Next.js 16 with Thirdweb for wallet auth and Recharts for charting NAV over time. We went for a dark Bloomberg-terminal look - the whole point is to show investors everything about performance while revealing nothing about the strategy.