Skip to content

About OccurAI

A practical first step toward photonic computing.

Our current project is a programmable optical interconnect built around a soliton microcomb. The first product combines a routing PIC, control software and an evaluation module.

Current project

Communication between compute modules.

A microcomb supplies several optical wavelength carriers. Transmitters put independent data on those carriers, and a programmable circuit routes them to receivers. Electronics coordinates path changes, calibration and receiver readiness.

We are targeting short links on custom accelerator carrier boards, with endpoint interfaces defined together with an evaluation partner. The first routing chip is intended to work with an existing or shared source.

Published work has demonstrated soliton-microcomb optical switching. Our development focus is a compact routing module, its control software and measured integration with compatible endpoints.

Related work: soliton optical circuit switching

Approach

Requirements, prototype, evaluation.

  1. Define the link. Agree the interfaces, traffic pattern, current alternative and improvement that would justify integration.
  2. Build the prototype. Secure source and laboratory access, complete the routing design with an experienced photonics lead, then fabricate and package the circuit.
  3. Measure the system. Evaluate bit errors, optical loss, crosstalk, command-to-valid-data time and complete-system power under the agreed conditions.

The project is currently at the design and feasibility stage. We are looking for technical collaborators and teams interested in defining an evaluation.

Research direction

Computing and topological representations.

The longer-term goal is a programmable photonic processor with electronic control. A photonic NPU is a possible later direction, informed by experience with sources, packaging, calibration and interfaces.

Our interest in topology began with using mathematical structure as a representation for machine learning. Earlier experiments compared neural-network architectures and topological features on protein point clouds. Those experiments remained inconclusive, with memory limits and representation choices constraining the evaluation.

Topological information encoding, including ideas involving Hopfions, remains an exploratory research direction. Defining useful writing, manipulation and readout operations is part of that work.

Earlier paper on topology for ML (PDF)
Pavel Chekmaryov

Founder

Pavel Chekmaryov

Software and ML background, with work on computational models, chip-test tools and exploratory photonic layouts. Previous work also includes building a startup platform and technical hiring.