Decentralized AI Inference Network
Sight AI is a decentralized AI inference network. It offers a unified API compatible with OpenAI, providing access to hundreds of proprietary and open-source AI models. Additionally, it allows individuals to become providers by contributing GPU, local, edge computing resources, or idle APIs.
What problems are we solving?
Unified multi-model access: Developers no longer need to integrate each model/vendor, billing scheme, and rate limiter one by one—use a single gateway and common protocol.
Reliability and low cost end-to-end: Traffic is automatically routed by price, availability, and latency, balancing stability with cost control.
Open supply side: Community members can contribute compute and upstream API keys. Zero-trust encryption and verifiable settlement ensure safety and fairness.
Key components
3.1 AI Gateway (Unified API, Billing, Scheduling)
OpenAI-style endpoints aggregate multiple closed/open models.
Auto-routing by price / availability / latency.
Unified balance, billing, and reconciliation.
The model you pass maps to the best provider; the Gateway completes the execution chain (Gateway → Executor → Upstream Model).
3.2 AIO (Adaptive Inference Orchestrator)
Node metrics: latency, success/failure, throughput; heartbeats (30/60s), multi-window observation.
Candidate filtering & capacity control: track idle/busy/offline, avoid overloads.
Tiered dispatch: prefer nodes outperforming the 5-minute average; otherwise fall back to threshold sets with retries.
3.3 Executors (Stateless adapters)
Perform the actual upstream calls (OpenAI/Anthropic/DeepSeek, etc.).
Register to etcd with /executors/{region}/{model_type}/{executor_id} and send heartbeats with TTL; periodically report latency and load.
The scheduler uses a combined cost (e.g., latency × load) to route.
3.4 BYOK / Share API Key (Supply marketplace)
Zero-trust encryption: Providers encrypt keys client-side using X25519 + ChaCha20-Poly1305. The Gateway stores only ciphertext; only the Executor (with the private key) can decrypt and use it.
Public-key and ciphertext structures support multi-region registration and revocation.
3.5 Settlement & Audit (Verifiable)
For each period, the platform produces a Snapshot (with Merkle root and price table) so everyone can locally verify usage-based fees and revenue sharing.
Claims are finalized via EIP-712 signatures after local validation. Exported logs and schemas enable reproducible checks.
Who it’s for & value proposition
Audience
What they care about
What we provide
AI App Developers
One-stop access, model aggregation, predictable cost
OpenAI-style API + auto-routing by price/availability/latency + full billing & reconciliation
Compute/Edge Owners
Low-friction monetization, transparent status
One-click Miner/Client; dashboard for jobs, earnings, device health (CPU/GPU/memory/uptime)
API Providers
Security and fair revenue sharing
Client-side encryption (no plaintext at rest), snapshot + Merkle verification, signed on-chain claims
Last updated