Skip to content
Toleron

Technology · Edge AI · Compute

Inference where the raw stream is.

A model that needs kilohertz features has to run where those features can be computed — at the machine. Toleron executes ONNX models on the EdgeBox or on a GPU-equipped AI box over the LAN, only on in-trigger data, with every model verified before it runs.

Runtime
ONNX, verified on host
AI box
Jetson Orin Nano class
LLM
Local, tool-calling

Placement

The planner decides where a pipeline runs by data locality, capability and budget, honours the user's preferred host when capabilities allow, and records the reason for every fallback. Capabilities are probed, not declared: a box counts as model-capable only after the runtime is verified on it. Proven live: an AI box computed one-second windows from a neighbouring EdgeBox's raw stream, with the results stored centrally and exactly one producer.

Verified artifacts

Models are validated at upload — optionally against reference samples — and verified again before every execution. Plugins that extend the operator set are signed, admitted per host with an audit record, can be withdrawn everywhere with one action, and run isolated with resource limits.

Local language models

The assistant and Toleron Flow run on local language models on the customer's own hardware, with native tool calling. Documents and operational data never leave the site.

Next step

Bring intelligence to your machines.

Connect existing CNC equipment without replacing the machine or the controller. We start with one machine, real data and an honest acceptance run.