Stays
The model, the image, and the case file live on the client's machines. Inference does not leave the premises.
Artificial intelligence
A public AI service does not come in when the document cannot leave. KAI runs the model on the premises, on a multi-agent platform over the CSF grid.
The model, the image, and the case file live on the client's machines. Inference does not leave the premises.
Text, vision, voice, or documents. The grid leader hands it the work.
The system submits the work and moves on. KAI returns a ticket, the tracking identifier, and the poll repeats until the answer is ready.
Why not a public service
The image, the document, and the voice are processed on the client's machines. A public artificial-intelligence service stays outside that perimeter.
Adding a machine means joining it to the CSF grid. The leader assigns the work. There is no load balancer and no peer list to configure.
Agents
KAI is a multi-agent platform. Each agent takes an objective. The leader assigns the node by quota and by whether the model is already on that machine.
RAG retrieves the material published on the premises. CAG, Context Aware Generation, generates the answer with the context of the job, not a loose answer from the model.
Objects in video, license plates, and vehicle tracking across cameras. Faces and people, also when the face is covered.
Whisper turns voice into text. Kokoro turns text into voice. Both run in Java, on the node.
The model runs in Java, in the same process that handles the case file. Tensors, networks, weights, and dictionaries stay on the premises.
How a system consumes it
The system submits the work and moves on. KAI returns a ticket. The poll repeats on that ticket until the answer is ready.
If the system has CSF, Crossover SOA Framework, installed, it publishes into a folder on the grid. The event is replicated on the KAI nodes.
A file or a batch enter through the same tracking identifier. Another core is not required.