Ideas before product.
How DeckerGUI began.
DeckerGUI started with a dissatisfaction: AI tools were being toyed with, not operated. Organisations adopted models the way children adopt pets — individually, chaotically, and with no shared system underneath.
So we built the layer everyone skipped — an orchestrating framework that lets personas, workflows and tools run as one governed machine, whether on the cloud, at the edge, or with the network pulled out.
Three research pillars, one thesis.
DGUIYOLOMoE
Vision-centric agentic routing.
Routing that starts from a vision of the goal — a mixture-of-experts choreographer that routes work to the right model, tool or persona, and validates the trajectory before it acts.
TCR
Conditional response.
A response model that conditions every AI answer on persistent, shared context — so agents stop hallucinating about the state of the world and start reporting on it.
JEPA
Joint-embedding predictive architecture.
Predictive learning that builds internal world-models by predicting embeddings — the perceptual core behind context-aware coordination.
Extended by CTAX-Ai and the Big Pickle reference runtime. See the full thesis in the DeckerGUI whitepaper.
What we admit is hard.
Context
Grounding agents in durable, shared state instead of isolated prompt memory.
Containment
Letting agents act without letting them loose — fail-closed gates and role boundaries.
Offline operation
Running intelligent coordination where the network stops — edge and air-gapped modes.
Metering
Making every token, tool call and agent action accountable through durable ledgers.
Research you can ship.
We publish what we build — and build what we publish.