MindSafe - Private AI Mental Health Screening

• 

  Main Page / Dashboard of Mindsafe

• 

  Live Conversation with Medical AI

• 

  1AM Wallet

• 

  Result (With Option to Mint or End Session)

Inspiration

1 in 4 people have a mental health condition. 75% never get help — not because care doesn't exist, but because getting screened creates a permanent record that can affect your insurance, your job, your life.

I kept thinking: what if getting help left absolutely no trace?

What I Built

MindSafe is a privacy-first mental health screening dApp on Midnight.

An adaptive AI clinician asks dynamic questions — not static forms — and adjusts based on your responses the way a real clinician would. The entire conversation lives only in browser memory. No server. No database. No localStorage. When the session ends, everything is gone.

At the end of the screening, a zero-knowledge proof is minted on the Midnight blockchain recording only the outcome — MILD, MODERATE, or SEVERE — and a cryptographic hash. Never the conversation. Never your identity.

$$\text{proof} = \text{ZK}(\text{score}) \rightarrow {\text{category}, \text{hash}}$$

The result is something that didn't exist before: a verifiable mental health screening that you can optionally share with a therapist or insurer, without revealing a single word you said.

What I Learned

Midnight's privacy model rewires how you think about data. In normal development, you design around what you store. On Midnight, you design around what you disclose — and only that. Writing the Compact contract made me realize how much unnecessary data most health apps collect simply because nobody asked "do we actually need this?"

ZK proof generation (~60–120 seconds per transaction) also taught me that real cryptographic privacy has a cost — and that cost has to be designed around, not hidden from the user.

Challenges

• Compact learning curve — almost no examples existed for health or AI use cases. Circuit errors gave one-line messages with multi-hour root causes.
• Witness wiring — getting the private severityScore witness to flow correctly from TypeScript into the ZK circuit without disclosure required deep SDK debugging.
• Proof server latency — the Docker proof server takes 60–120s per proof. The 1AM wallet confirms the transaction, but the indexer subscription sometimes never resolves, leaving the UI stuck on "minting." I built graceful fallbacks, but it remains the biggest UX gap.
• SDK gaps — midnight-js v4 has breaking changes from earlier versions with limited migration guidance. I ended up reading source code and MCP tool outputs to find the correct patterns.

Built With

• 1am-wallet
• compact-(midnight-zk-language)
• docker
• gemini-api-(gemini-2.5-flash)
• midnight-js-sdk
• midnight-network
• react
• typescript
• vite
• wasm
• wsl2