TheFoundationalModelforRadioNetworks
The RA-NN — Radio Access Neural Network — redesigns the RAN as a learnable system. Not AI on top. AI as the architecture.
AI-RAN has an architecture problem
Every AI model solves one problem
PCI planning. Mobility optimization. Load balancing. Each gets its own model, its own data pipeline, its own view of the network. None of them share what they learn.
The network is observed through averages
Aggregate KPIs collapse thousands of individual user experiences into a handful of counters. The real behavior — the spatial variation, the tail cases, the localized failures — disappears.
Intelligence is bolted on, not built in
AI gets added on top of architectures designed decades ago. The underlying abstractions were never meant to learn. No amount of ML on top changes that.
So we built a new one
Radio networks are physically constrained — propagation, interference, mobility, spatial locality. You can't decompose them into independent problems and expect coherent behavior. The RA-NN treats the entire network as one system that learns its own structure.
Observe at the user level
The RA-NN works from individual user measurements — the actual physical signals at the edge — not the compressed summaries that legacy systems rely on.
One representation, all tasks
Coverage, interference, mobility, load balancing — all derive from the same learned model of the network. Solve one problem and the next one starts ahead, not from scratch.
Structure is learned, not configured
The inter-cell relationships are inferred directly from what the network observes. No predefined topology. No manually maintained neighbor lists.
Operator DNA
Decades inside Tier 1 networks
Our founders spent years operating and optimizing the networks the RA-NN is designed for. The architecture didn't come from a research lab — it came from seeing what actually breaks when you try to make these systems intelligent.
PCI optimization in seconds — not hours
Applied to Physical Cell Identity planning, the RA-NN learns the network's interaction structure from live observations — turning a combinatorial search that takes minutes or hours into a direct inference in seconds. The same structure generalizes to mobility, load balancing, and interference without starting over.
Spatial reasoning from RF alone
Given only received signal measurements and cell-level RF characteristics, the RA-NN learns the spatial geometry of the radio environment — resolving where users sit within surrounding antenna beams without explicit propagation models or timing data. Understanding where a user is in the RF environment is the prerequisite for reasoning about how to improve it. Accuracy scales with the number of observed cells, confirming the model is learning structure, not memorizing patterns.
Early-stage results — active developmentThe architecture exists. Now we need the right partners
We're looking for operators and infrastructure teams who see the limits of the current approach and want to build on the next one.