ZK_COACH

Most AI coaches have two fatal flaws: they agree with everything you say (sycophancy), and they forget everything when the session ends (ephemeral memory). ZK Coach fixes both. It provides a multi-persona advisory board—featuring roles like the Validator, the Challenger, and the Realist—that actively pushes back on your rationalizations by cross-referencing your current statements against your historical behavioral patterns.

To do this securely, ZK Coach leverages decentralized infrastructure. Your highly sensitive coaching history is stored as encrypted, user-sovereign blobs on Walrus, guaranteeing you control your data. We use World ID for proof-of-personhood, giving verified humans Sybil-resistant trial access to the AI coaching team.

Finally, ZK Coach introduces real financial accountability via Hedera. Users can set a goal and stake real HBAR on it. The commitment is immutably logged, and if the AI evaluates that you failed to meet your goal, your stake is automatically transferred to charity. It’s "therapy with teeth."

ZK Coach is built with a Node.js/Hono backend and a custom orchestration engine that manages multiple LLM personas simultaneously.

We integrated three core partner technologies to power the product:

1. Walrus (Sui): We built a custom walrus-store package using the @mysten/walrus SDK to completely replace a traditional database. Every coaching session and parsed behavioral pattern is serialized and written as an encrypted blob to the Walrus testnet. The AI retrieves these blobs to construct its context window, giving the user portable, sovereign ownership over their most sensitive mental health data.
2. Hedera: We implemented the @hashgraph/sdk to build our "Commitment Staking" engine. When a user commits to a goal, they stake HBAR, and the goal parameters are immutably logged via the Hedera Consensus Service (HCS). On the deadline, the AI evaluates the outcome and triggers an automated HBAR transfer—returning the funds to the user on success, or routing them to a charity on failure.
3. World ID: We integrated @worldcoin/idkit-core to gate the application. To prevent abuse of the AI advisory board, World ID verification serves as an anti-Sybil lock, granting real humans an initial free trial of coaching sessions.

Notable Hack / Architecture Choice: Instead of relying on standard vector database RAG (which can be fuzzy and lossy), we developed a structured markdown "wiki" system. The AI continuously synthesizes and writes deterministic facts and patterns back to the user's decentralized Walrus storage. This explicit read-write loop ensures the "Challenger" persona has exact, undeniable quotes from past sessions injected directly into its prompt, allowing it to forcefully (but politely) push back when the user contradicts themselves.