Keepsecret

• 

Inspiration

We live in the era of the "Digital Trail." Every time we ask an AI for medical advice or financial planning, we trade a piece of our soul for a bit of intelligence. But there is a silent predator lurking: Harvest Now, Decrypt Later (HNDL). Malicious actors are collecting encrypted data today, waiting for the quantum computers of tomorrow to break the locks. We wanted to build a world where you don't have to choose between an Amnesic AI (that forgets you) and an Invasive AI (that exposes you). We wanted to build a "Brain in a Vault."

What it does

Keepsecret is a "Stateless AI Agent" that provides high-utility personal assistance without ever "owning" your identity.

The Awakening: When you log in with your Abelian wallet, the AI "rehydrates" by retrieving your encrypted history from the blockchain.

The Interaction: You talk to a powerful DeepSeek-R1 model that knows your preferences, allergies, and goals.

The Sleep: When you log out, the AI distills the conversation into a tiny, high-density memory fragment, encrypts it with Post-Quantum Cryptography (PQC), and locks it back onto the Abelian blockchain.

The Result: The local device is wiped. The AI provider is wiped. Only you hold the key to your memory.

How we built it

We integrated three "frontier" technologies:

DeepSeek-R1: Acts as the "Distiller," using advanced reasoning to compress complex dialogues into 512-byte JSON fragments.

Abelian Network: Serves as the quantum-resistant storage layer. We used the Abelian SDK to implement lattice-based encryption, specifically ensuring the ciphertext C is secure under the Learning With Errors (LWE) assumption.

Python Sovereign Client: A local environment that manages "Ephemeral RAM," ensuring no plaintext logs ever touch the physical disk.

Challenges we ran into

The hardest part wasn't just the code; it was the human element. Our team is a "melting pot" of different majors—from CS and Cybersecurity to Finance and Design.

The "Jargon Gap": The developers wanted to talk about ring signatures, while the finance majors were worried about on-chain gas costs, and the designers were worried about user friction.

Concept Friction: We had intense debates on "AI Autonomy." Should the AI decide what is a "secret," or should the user?

Technical Constraints: Fitting a personality and context into a <1KB blockchain payload felt like trying to fit an ocean into a glass. We had to iterate on our "Distillation Prompts" dozens of times before the memory felt "human" yet efficient.

Accomplishments that we're proud of

The "Invisible" PQC: We successfully integrated Abelian’s complex lattice-based cryptography into a seamless flow where the user doesn't even feel the "quantum-resistant" armor they are wearing.

Semantic Density: We achieved a 95% compression rate on chat history using DeepSeek-R1 without losing the "emotional core" of the user's context.

Proof of Erasure: Demonstrating that after a session, even with forensic tools, zero data could be recovered from the local machine.

What we learned

We learned that Privacy is a Team Sport. Without our finance teammate, we would have built an expensive, unmarketable tool. Without our designers, we would have built a tool too complex for anyone to use. We realized that solving the "Trust Gap" in AI requires a multi-disciplinary lens—combining game theory, cryptography, and linguistic psychology.

What's next for Keepsecret

QDay Expansion: Migrating "Hot Memory" to Abelian’s L2 (QDay) to allow for real-time memory syncing with near-zero fees.

Encrypted RAG: Developing a way for the AI to "search" through a library of encrypted Abelian blocks without decrypting the entire history at once.

The "Dead Man's Switch": Using smart contracts to allow specific "Context Fragments" (like a medical history) to be released to a trusted party in case of emergency—all without compromising the user's root privacy.

Built With

• abel
• aes-256-gcm
• cli
• deepseek-r1
• ml-kem
• pynacl
• python
• qday