AutoMLApp (in development)
Design ML pipelines on a canvas. Run them on your hardware.
A Windows node-graph editor for composing, validating and running machine-learning and accelerator pipelines.
The Visual Node-Graph Workbench
AutoMLApp is the visual node-graph workbench for heterogeneous compute and ML pipelines. Instead of wiring configuration files by hand or writing fragile glue code, you drag nodes onto a canvas, connect typed ports, and compile the graph into an execution plan.
It unifies data pipelines, model deployment, accelerator configuration, real-time telemetry and hyperparameter search in one comprehensive, high-productivity desktop workspace.
Real-time visual link rejection prevents invalid data bindings as you draw connections.
Compiles graph geometry into optimal step execution orders with domain handlers.
HyperSwarm balances raw model accuracy directly against hardware compute costs.
185 Node Types Across 25+ Specialized Domains
End-to-End Visual Engineering Features
Engineered for high-assurance ML workflows, hardware accelerators, and deterministic agent orchestration.
Node-graph editor
185+ Node Types • 25+ DomainsDrag and drop across 25+ domains including data pipelines, ONNX, IPU, spiking networks, swarm, FPGA, RDMA, feature store, and sidecar. Typed connectors reject incompatible links in real time as you draw them.
Execution engine
Topological Sorting & Domain HandlersCanvas graphs compile automatically into topologically sorted execution plans with domain-specific step handlers dispatched directly to host accelerators or remote clusters.
Instant validation (F7)
Comprehensive Invariant CheckingA single keypress (F7) runs static analysis checking for cycles, missing dependencies, duplicate IDs, orphaned nodes, and port type mismatches, immediately flagging defects on canvas.
Integration bridges
28 Native Subsystem ConnectorsAround 28 specialized bridges connect canvas nodes directly to HyperSwarm, ONNX runtime, spiking neural networks, SwarmDev deterministic pipelines, the telemetry sidecar, and hardware engines.
HyperSwarm-backed AutoML
Cost & Score-Aware Architecture SearchHyperparameter search executes on HyperSwarm through a pluggable backend where candidates define model family, depth, learning rate, batch size, and dataset. A telemetry-aware engine balances accuracy against compute cost signals.
Workspaces & panels
40+ Subcanvases & Specialized ToolingExperiment, admin, and data-catalog zones with 40+ subcanvas types and dedicated side panels for hyperparameters, IPU tile connections, cluster setups, deployments, observability, model registries, and test runners.
Portable pipelines
Two-Way JSON InteropExport and import standardized pipeline JSON schemas fully compatible with the web pipeline manager, background daemons, and command-line execution runners.
Editing fundamentals
Productive Desktop ErgonomicsFull undo/redo stack, automatic force-directed graph layout, project state persistence, execution run histories, and diffable step checkpoints.
Tiered entitlements
Granular Feature AccessNode availability, accelerator targets, and advanced swarm algorithms can be dynamically filtered and partitioned by organization license tier.
Runtime Status & Subsystem Maturity
Status: Windows desktop app (.NET 10, WinForms). The pipeline editor, validation (F7), and HyperSwarm backend are working. Stage-by-stage run progression and tile telemetry are currently simulated, and the telemetry-first AutoML engine is a proof of concept.
Pipeline Canvas & Graph Editor
185 node types, port connection type checking, drag-and-drop wiring, and serialization into standard JSON pipelines.
Validation Engine (F7) & HyperSwarm
Cycle detection, orphaned node tagging, and pluggable HyperSwarm backend candidate execution (family, depth, learning rate, batch size, dataset).
Telemetry & Stage Progression
Stage-by-stage run progression and hardware tile telemetry are simulated. The cost-weighted telemetry-first AutoML engine operates as an active PoC.
{
"$schema": "https://hflabs.dev/schemas/pipeline_schema.json",
"pipelineId": "pip_hybrid_accelerator_swarm_01",
"version": "1.0.0",
"targetRuntime": "HFLabs.NET10",
"nodes": [
{
"id": "node_01",
"type": "Ingestion.FeatureStoreRef",
"domain": "DataPipelines",
"config": { "dataset": "kge_triples_v2", "batchSize": 256 },
"outputs": [{ "port": "out", "type": "TensorBatch" }]
},
{
"id": "node_02",
"type": "SwarmDev.IntentRouterNode",
"domain": "Cognition",
"config": { "schemaVersion": "v1.2", "policyGating": "Strict" },
"outputs": [{ "port": "plan", "type": "SubGraph" }]
},
{
"id": "node_03",
"type": "AutoML.HyperSwarmSearchNode",
"domain": "SwarmOptimization",
"config": { "evaluator": "CostAwarePareto", "subSwarms": 4 },
"inputs": [
{ "port": "data", "sourceNode": "node_01", "sourcePort": "out" },
{ "port": "plan", "sourceNode": "node_02", "sourcePort": "plan" }
]
}
]
}Explore the HFLabs Visual & Compute Stack
Discover how AutoMLApp orchestrates Hypertensor GraphRAG, SwarmDev pipelines, and heterogeneous hardware accelerators.