Sealed-Open Project Idea

[joycechinnock-pixel]

SEALED — PROJECT PITCH
for SIG Community consideration

QUICK RECAP

Sealed is a protocol that lets organizations with sensitive data (hospitals, banks, law firms, research labs) get AI trained on that data without the raw data ever being exposed to anyone — not the model provider, not the compute network, not Sealed itself. It stitches together three existing pieces of infrastructure that have each matured on their own but have not been combined this way before.

ARCHITECTURE — WHAT WOULD ACTUALLY BE USED

This is the real component mapping, not a vague concept — naming the exact pieces this would run on if built:

1. Data Owner encrypts the raw data client-side using Zama's TFHE-rs / Concrete ML (fully homomorphic encryption libraries) → produces ciphertext. The decryption key never leaves the Data Owner's side.

2. Ciphertext is deployed as a compute workload via an Akash SDL (Stack Definition Language) manifest → matched through the Akash Provider Marketplace to a GPU-equipped Provider → leased via an Akash Deployment, at a fraction of AWS/Azure GPU pricing.

3. A dedicated Bittensor Subnet ("Sealed Subnet") handles the AI work itself: Miners train/run models directly on the ciphertext using Zama's fhEVM-compatible execution, never decrypting it; Validators score the quality of the encrypted output through Bittensor's Yuma Consensus, earning TAO emissions for accurate scoring.

4. Encrypted datasets and model checkpoints are persisted via Filecoin storage deals rather than centralized cloud storage.

5. The encrypted result is returned to the Data Owner, who is the only party holding the decryption key, and decrypts locally.

Flow: Data Owner → [TFHE-rs/Concrete ML encryption] → Akash SDL deployment → Akash Provider Marketplace/GPU lease → Bittensor Sealed Subnet (Miners train, Validators score via Yuma Consensus) → Filecoin storage → back to Data Owner → local decryption.

WHY NOW

Each piece above already exists and is live today. The gap is that nobody has wired them together specifically for regulated data — that's the opening.

HOW THIS DIFFERS FROM TEE-BASED CONFIDENTIAL COMPUTE PROPOSALS

Akash has already seen proposals for confidential AI execution built on Trusted Execution Environments (TEEs — e.g. the TEE-based confidential compute SDK proposal in discussion #872). That approach is hardware-trust-based: you're trusting Intel/NVIDIA's chip-level isolation and attestation not to be compromised or backdoored, and TEEs have a real history of side-channel exploits. FHE is a different, stronger guarantee: the data is never in a decrypted state anywhere in the pipeline, on any hardware, so there's nothing for a compromised provider or a side-channel attack to expose. The tradeoff is speed — FHE is slower than TEE execution today — which is why this is scoped to narrow high-value tasks rather than general compute.

FIRST MILESTONE (deliberately narrow)

Rather than the full five-part pipeline at once: a single working demo of one use case end-to-end (e.g. encrypted fraud-detection scoring) on the Bittensor Sealed Subnet + Akash compute layer, before expanding to other verticals. Scoped small on purpose.

HONEST LIMITATION

FHE computation is still slow. This is a "regulated industries, narrow high-value tasks" play first (healthcare diagnostics AI, legal-tech document analysis, financial fraud detection), not a mass-market or frontier-LLM-training play yet.

TARGET VERTICALS

Primary: healthcare, legal, financial services (fraud detection).
Also under consideration: fire brigades/emergency services, agricultural chemical businesses, and other crypto/Web3 projects as early technical partners.

PURPOSE OF THIS POST

Not a funding request. Posting here for SIG Community feedback on the design, and to find anyone in the ecosystem building on Akash + FHE + decentralized compute who might want to collaborate. Funding (if any) is being pursued separately, outside Akash's own pool, thats not to say suggestions or offers will be declined either though, all useful input is welcomed.

TEAM

Two-person team, already collaborating on other projects together: Tristan (backend) and Jocelyn (frontend) — a working split, not a rigid boundary. Open to further co-builders from the Zama, Bittensor, or Akash ecosystems.

Thank you for your time and were looking forward to hearing your thoughts!

REFERENCES

Zama — TFHE-rs / Concrete ML / fhEVM: https://github.com/zama-ai
Bittensor subnet mechanism (Yuma Consensus): https://bittensor.com
Akash SDL / Provider Marketplace: https://akash.network
Filecoin storage deals: https://filecoin.io