The sensory nervous system for AI. Real-time intelligence on the hardware you already own.
Ulysses is a real-time sensory AI system built on Energy-Based Models, co-designed with Intel Labs. It processes raw, streaming, physical-world data at sub-10ns latency on standard Xeon CPUs — no GPU required.
Ulysses sits between raw reality and every AI system downstream. It processes, detects, and learns from streaming data — then feeds clean intelligence to everything else.
Streaming physical-world data arriving in real time from diverse sources.
The sensory layer. Processes, detects, and learns from streaming data in real time.
Every model, pipeline, and system that needs to understand the physical world.
The data pipeline is simple: Raw Reality → Ulysses → Everything Else. Ulysses doesn't replace your AI stack. It makes every part of it better by giving it real-time sensory intelligence.
Transformers are brilliant at words. They're terrible at streaming physical-world data. AI needs both halves of the brain — and right now, one half is missing.
Transformers and LLMs excel at language, reasoning, and structured knowledge. A $6 trillion ecosystem of proven capability.
Energy-Based Models process raw, streaming, numerical data from the physical world. This is where Ulysses operates — the missing half of intelligent AI.
Ulysses doesn't compete with transformers. It completes them. The human brain has both a cognitive system and a sensory system. Current AI only has the cognitive half. Ulysses is the sensory system that feeds everything else.
Not a better model on commodity infrastructure. A system where model and hardware execution are inseparable.
Model and hardware co-designed as an inseparable system with Intel Labs. Zero-copy active inference architecture eliminates data movement overhead entirely.
No pretraining required. Ulysses learns from streaming data as it arrives, continuously adapting to new patterns and regime changes without retraining cycles.
Runs on Intel Xeon CPUs with AMX acceleration. Deploys where your data lives — at the edge, on existing hardware, with no specialized accelerators required.
34x fewer parameters than comparable architectures. Energy-Based Models achieve superior performance with radically less computational overhead.
Inference at the speed of hardware, not software. The zero-copy architecture delivers results in under 10 nanoseconds — fast enough for any real-time application.
Acts as the sensory front-end for any downstream system. Feeds LLMs, ML pipelines, analytics platforms, and domain-specific models with real-time intelligence.
Benchmarked against comparable architectures on real-world detection tasks. These aren't theoretical — they're measured.
34 times fewer parameters than comparable architectures while achieving superior detection accuracy.
100% regime recall on COVID detection benchmark. Standard models achieved 26% on the same task.
Sub-10 nanosecond inference. Hardware-speed execution through zero-copy active inference architecture.
Runs entirely on Intel Xeon CPUs with AMX. Deploys on existing infrastructure with zero additional hardware cost.
From vertical applications to horizontal platform. The roadmap from product to infrastructure.
Ulysses delivers immediate value through targeted applications in signal processing and anomaly detection across specific domains.
Ulysses becomes the sensory front-end for all hybrid AI architectures — a universal layer between the physical world and every downstream AI system.
Optimized for Xeon with AMX. The model and hardware aren't just compatible — they're co-designed as an inseparable execution system.
Most AI companies build models and then find hardware to run them. Kavara did the opposite. Ulysses was co-designed with Intel Labs from the ground up, creating a zero-copy active inference architecture that treats the model and the silicon as a single system.
The result is inference at the speed of hardware, not software. No data movement. No memory copies. No GPU tax. Just raw execution on the billions of AI-capable transistors sitting idle in existing infrastructure — what we call dark flops.
Every Intel Xeon processor with AMX already has the capability to run Ulysses. That means the hardware is already deployed. It's already paid for. It's just waiting.
Native execution on Advanced Matrix Extensions without GPU dependencies.
No data movement overhead. Model execution fused with hardware memory paths.
Billions of AI-capable transistors in existing infrastructure, activated.
Ulysses deploys on the infrastructure you already own. No GPUs to procure. No cloud dependencies. Real-time intelligence, running today.