P2P AI inference that heals itself. Gossip-replicated ledger, automatic task recovery.
Whisper Network is a coordinator-free AI inference marketplace built on Gensyn's AXL encrypted P2P mesh.
Every node holds a full replica of the distributed task ledger. When a query arrives, it's auctioned across the network—the cheapest provider wins and executes. If that node dies mid-execution, survivors detect the failure within seconds, reclaim the orphaned lease, and a backup node completes the work automatically.
The system routes queries to specialized agents: ML agents handle neural networks and transformers, Web3 agents specialize in smart contracts and DeFi, DevOps agents handle infrastructure queries. Each agent has a domain-specific system prompt that augments the query before sending it to Ollama for real LLM inference.
No central broker. No third party. Just nodes, gossip, and agents. Kill 50% of the nodes mid-execution—the system still completes all tasks.
Architecture:
Layer 1 (Membership): SWIM-lite failure detection. Heartbeats every 2s. Nodes suspected dead after 6s silence, confirmed after 2+ peer reports. Fast-suspect via AXL topology drops.
Layer 2 (Ledger): Append-only task log replicated via hop-limited gossip. Vector clocks resolve conflicts. Completed tasks are sticky (cannot regress). Persisted to JSON for crash recovery.
Layer 3 (Agents + Runtime): AgentRegistry with 3 agents per node (ml-agent, web3-agent, devops-agent). Keyword-based routing finds best agent. Each agent has specialized system prompt. Polling loop claims expired leases, executes tasks, gossips results.
Layer 4 (Inference + ENS): Real Ollama LLM calls (fallback to keyword search). PyENS library registers node/agent names on Sepolia. Results include agent_ens_name for UI display.
Tech Stack:

