Pathway's machine unlearning keeps shipping ETAs honest as data streams go stale

French deep-tech startup Pathway, co-founded by CEO Zuzanna Stamirowska — author of the state-of-the-art maritime trade forecasting model published by the National Academy of Sciences of the USA — attacked a problem slowing logistics digitization: the lack of software infrastructure that can do automated reasoning on top of real-time data streams. Sensors on shipping containers generate tracking data, but low-bandwidth environments delay cloud syncing, producing a mismatched deluge of delayed and live data.

The framework retrains AI systems and LLMs continuously on streaming data, making it possible to revise individual data points without a new full batch upload — machine unlearning, the challenge Google launched a competition on earlier that summer. With shipping line CMA CGM, Pathway improved the precision of container gate-out ETAs and improved terminal operations, speeding up container handling times and lowering business and environmental costs.

The company stresses the approach is industry-agnostic across data types — table-like data, time series, IoT messages, event streams, things in motion, graphs and ontologies — and says it is focused on product refinement, including observability use cases.

Stale or wrong training data silently corrupts model outputs in businesses that run on streaming data.

Full batch retraining is too slow and costly to run every time one fact changes.

Logistics decisions depend on real-time data, so corrections must land while they still matter.

In fast-moving data, the unit of work isn't the dataset but the single fact: design systems that forget and correct continuously.

Tech.eu reported the CMA CGM deployment as traction in industries reliant on accurate real-time data, while noting there is no single solution to unlearning — batch, streaming and LLM cases all differ. Pathway stayed focused on refining the platform for multiple data types and industries.