United Nations

Winner

Innovate for Impact Award
AI for Good
Security Architecture

Security by
Architecture

No single server ever sees your complete inference — every connection is encrypted and every handoff is cryptographically signed.

TLS 1.3 Encrypted
Ed25519 Identity
HMAC-SHA256 Signed
Input Validated
Security architecture

Four Layers of Protection

Every request passes through four independent security layers.

End-to-End Encrypted Transport

libp2p + TLS 1.3

Every connection is independently encrypted with TLS 1.3 so no intermediary can read data in transit.

Mutual TLS 1.3 on every connection
Independent encryption per hop
Bank-grade encryption standard

Cryptographic Server Identity

Ed25519 Peer IDs

Every server has a unique Ed25519 key pair, letting you verify its identity and preventing impersonation.

Unique Ed25519 key pair per server
Network poison-resistant
Unforgeable reputation chain

Activation Integrity Verification

HMAC-SHA256 Signing

Each server signs its activation tensors with HMAC-SHA256 so tampering is detected immediately.

Unique random key per session
Signature on every handoff
Tamper detection is immediate

Input Validation & Protection

Resource Boundary Defense

Every incoming tensor is validated against size bounds so malformed or oversized inputs never reach the model.

Size-bounded tensor validation
DoS and memory protection
Malformed input rejection

The Journey of a Secure Request

Follow a prompt from your device through the distributed network — encrypted and verified at every step.

1

Your Prompt

Plaintext on your device only

2

TLS 1.3 Encryption

Encrypted before leaving

3

Tokenization

Converted to vectors

4

Layer Splitting

Split across servers

5

HMAC Signing

Signed at each handoff

6

Verified Output

Integrity confirmed

Proven Cryptographic Standards

Battle-tested protocols trusted by the world's most security-critical systems.

TLS 1.3

Transport Layer Security

What it does

Encrypts all data in transit between your device and every server in the pipeline.

How it works

Mutual authentication with forward secrecy — even compromised keys cannot decrypt past sessions.

Also used by

Banks, governments, healthcare systems, military communications.

Ed25519

Edwards-curve Digital Signature Algorithm

What it does

Generates unforgeable cryptographic identities for every server node.

How it works

Each server’s public key serves as its PeerID; messages signed with the private key are verifiable by anyone.

Also used by

SSH keys, Signal Protocol, Tor network, cryptocurrency wallets.

HMAC-SHA256

Hash-based Message Authentication Code

What it does

Signs every tensor handoff between servers to detect any tampering.

How it works

A unique session key generates a tag for each payload; the receiver recomputes and compares to detect any alteration.

Also used by

API authentication, JWT tokens, AWS request signing, blockchain verification.

Distributed = More Private

No single server ever sees your complete inference — distribution is inherently more private than centralization.

The Relay Race Model
Your Prompt
L0-9
Server A
L10-19
Server B
L20-29
Server C
Verified Result
Server A
Sees only layers 0-9
Cannot reconstruct prompt
Server B
Sees only layers 10-19
Cannot reconstruct prompt
Server C
Sees only layers 20-29
Cannot reconstruct prompt

No Single Point of Visibility

Prompts become high-dimensional vectors before reaching any transformer block — no server ever sees human-readable text.

Layer-Level Isolation

Each server processes only a slice of the model's layers — none has enough context to reconstruct your prompt.

Signed Handoffs

Like a relay race — each server carries the baton for one leg, and every handoff is cryptographically signed.

Centralized vs. Distributed

Why distributing computation provides inherently stronger privacy.

Centralized AI

Single Server Architecture
One server sees your complete prompt
One server processes all model layers
Single point of compromise = full data exposure
Provider has complete access to your data
Trust is placed in a single entity

Distributed AI Compute

Multi-Server Architecture
No server sees your complete prompt
Model layers split across many servers
Compromising one server reveals only activation tensors
Cryptographic verification at every handoff
Trust is distributed and mathematically enforced

Ready to Run AI Securely?

Enterprise-grade security built into the architecture — not bolted on as an afterthought.

View Architecture
Network connection failed. Please check your internet connection and Supabase configuration.