Public Intelligence
System Architecture

Architecture

The technical model and system boundaries for distributed AI infrastructure.

Overall System Topology

High-level view of the three-plane architecture: the website portal for user onboarding, the scheduler core as the control plane, and the contributed node network as the execution plane.

Overall System TopologyThree-node system architecture: Website Portal entry point, Scheduler Core control plane, and Node Network execution plane. Connections show data flow from website to scheduler to compute nodes.[WEBSITE_PORTAL]Entry & Onboard[SCHEDULER_CORE]Control Plane[NODE_NETWORK]Execution Plane

Request Flow

End-to-end request lifecycle from client submission through scheduler routing to node execution. The active path shows successful routing with direct result return bypassing the scheduler for verification.

Request FlowRequest lifecycle from client submission through scheduler routing to node execution. Active path highlighted showing successful routing to EU-West node with direct result return path.[CLIENT_INTERFACE]Status: SUBMITTING[SCHEDULER_CORE]Status: ROUTING[NODE_EU_WEST]Status: EXECUTINGDIRECT_RESULT_PAYLOAD (verification_loop)

Geographic Scheduling

Latency-based node selection across global regions. The scheduler evaluates candidate nodes and routes to the minimum-latency target (EU-West at 12ms in this example).

Geographic SchedulingScheduler routing decision showing three candidate nodes with latency measurements. EU-West selected as minimum latency (12ms) vs US-East (78ms) and AP-South (284ms).[SCHEDULER_CORE]Latency Engine[NODE_US_EAST]Latency: 78ms[NODE_EU_WEST]Latency: 12ms (MIN)[NODE_AP_SOUTH]Latency: 284ms

Distributed Inference

Model layer partitioning across multiple compute shards for large-scale inference. Each shard processes a contiguous layer range with pipelined tensor passing.

Distributed Inference PipelineModel layer partitioning across three compute shards. Scheduler partitions layers 1-12 to Shard A, 13-24 to Shard B, 25-36 to Shard C. Pipeline flow shown with active connections.[SCHEDULER_CORE]Layer Partitioner[NODE_SHARD_A]Layers: 1 - 12[NODE_SHARD_B]Layers: 13 - 24[NODE_SHARD_C]Layers: 25 - 36

Future Platform Stack

Four-layer architecture vision: application APIs, orchestration coordination, decentralized transport, and local hardware runtime. The coordination layer is the active development focus.

Future Platform Stack ArchitectureFour-layer platform stack: Application API layer, Orchestration & Coordination layer (active), Decentralized Transport layer, and Local Hardware Execution Runtime layer. Each layer spans full width with descriptive details.APPLICATION / API LAYERClient SDKs, telemetry portals, developer API endpointsORCHESTRATION & COORDINATION LAYERGeographic scheduling engine, node registries, queue modelsDECENTRALIZED TRANSPORT LAYERP2P overlay routing, message heartbeats, secure gRPC tunnelsLOCAL HARDWARE EXECUTION RUNTIMECUDA/ROCm device engines, local resource benchmarks, sandboxes