Oneirix is building the complete stack for verifiable self-improving AI.
Powered by our state-of-the-art photonic chip.
Our patented topological design eliminates the manufacturing defects that have held back photonic computing for decades. The result: chips that work straight from the foundry, dramatically reducing cost and time-to-market.
Training a single large AI model can consume as much electricity as 100 homes use in a year. GPUs are hitting a wall. We're building the alternative.
We compute using photons — particles of light — which move data with almost zero heat loss. No cooling required.
Matrix operations happen at the speed of light — millions of times faster than electronic processors.
A single P-TPU replaces racks of GPUs while using less power than a laptop charger.
True artificial intelligence requires more than raw compute. It needs the ability to improve itself, verify its own reasoning, and remain under human control. We're building all three layers.
The AI That Knows When It Doesn't Know
Metacognition for machines. RSSI-C gives AI the ability to think about its own thinking — detecting hallucinations before they happen. Using complex-valued neural networks, we measure phase coherence across an ensemble of models. When neurons fire in sync, the output is reliable. When they diverge, the AI knows to say "I'm not sure."
Unlimited Context AI Orchestration
The brain that never stops learning. ONEIROS maintains a World Model of the environment, coordinates tasks through Hierarchical Task Networks, and operates in biological wake/sleep cycles. During sleep, it simulates improvement trajectories through a Branching World Model — then presents proposed changes to humans via the Dream Log.
Photonic Tensor Processing Unit
The world's first self-improving photonic processor. Weights are stored as physical crystallization states in Sb₂Se₃ phase-change material. Learning happens through thermodynamic energy minimization. A "Dream Log" interface ensures no self-modification occurs without human approval.
P-TPU computes at the speed of light through photonic interference, preserving phase information.
ONEIROS orchestrates task decomposition and routes to specialized models via HTN planning.
RSSI-C validates every output by measuring phase coherence — catching hallucinations natively.
Experience logged to buffer for later consolidation during the sleep cycle.
Experience Replay — Consolidate wake-phase experiences through equilibrium propagation on P-TPU.
Branching Simulation — ONEIROS generates hypotheses and simulates N improvement trajectories M steps ahead.
RSSI-C Validation — Check coherence, intent alignment, and safety constraints. Reject unsafe proposals.
Dream Log — Present validated proposals to humans. No modification executes without explicit approval.
From validated design to the AI infrastructure of tomorrow.
Architecture specification v2.1 validated. GDSII layout DRC-clean. SVD Compiler achieving machine-precision accuracy.
PDK integration with CORNERSTONE/IMEC. First MPW run submission. UK Patent filing completed.
Receive fabricated chips. Optical characterization and validation. Demonstrate MNIST inference on real hardware.
Second-generation P-TPU with 256×256 channels. Partner integrations for data center deployment.
Full photonic AI coprocessor ecosystem. In-situ training capability. The new standard for sustainable AI infrastructure.
All parameters verified through rigorous simulation and DRC validation.
Channels per layer
Phase shifters
Non-volatile weights
Weight resolution
Total optical loss
No active cooling
Residual error
Rules passed ✓
We're seeking strategic partners and investors to bring Oneirix's verifiable self-improving AI from validated design to first silicon.