Company Dossier · Live AI · Est. 2020
Pathway
The team building AI that thinks and learns in real time - because a model frozen at training is just a smart intern on day one.
The team building AI that thinks and learns in real time - because a model frozen at training is just a smart intern on day one.
Somewhere a Formula 1 strategist is watching a wall of telemetry move faster than a human can read it. Somewhere else, a logistics planner at La Poste is deciding where ten thousand parcels go next. And in a quieter room with worse coffee, a NATO analyst is trying to make sense of data that refuses to sit still. Three very different people, one identical problem: the world changed a second ago, and their software hasn't caught up.
This is the moment Pathway was built for. Not the demo, not the keynote - the second between "something happened" and "we know about it." Most software treats that gap as acceptable. Pathway treats it as the whole job.
The company makes infrastructure for live AI: systems that ingest data as it arrives and react with low latency, instead of waking up each morning with yesterday's knowledge. A large language model, as CEO Zuzanna Stamirowska likes to put it, "acts like a smart intern on day one." Brilliant, eager, and utterly unaware of anything that happened after its training cut-off. Pathway's answer is less a smarter intern than a nervous system.
At its core, Pathway is a source-available Python framework for ETL, stream processing, real-time analytics, and the AI pipelines - RAG, LLM apps, vector search - that everyone now wants and few can keep current. The clever bit is unglamorous and exactly right: the same code runs on a historical batch and on a live stream. Write once, and the pipeline keeps itself in sync with SharePoint, Google Drive, S3, Kafka, PostgreSQL, and real-time APIs, no rewrite required.
It is Docker-friendly, Kubernetes-ready, and deployable on-premise for the kind of customer whose data never leaves the building. Underneath sits a high-throughput, low-latency engine that the company bills as one of the fastest on the market - the part that lets a Formula 1 team chew through telemetry at reportedly 90x prior speeds.
Source-available Python ETL for stream processing, real-time analytics, LLM pipelines and RAG. Same code, batch or live.
Docker-friendly, ready-to-run RAG and enterprise search that stays always in sync with your data sources.
The fast engine packaged for production: on-prem or cloud-native, Kubernetes-compatible, enterprise-grade security.
A post-transformer model with linear attention and unlimited context, aimed at continual learning and long-horizon reasoning.
Reported figures from named deployments. Speed and cost are two sides of the same coin: when the engine is fast, the bill gets smaller.
Complex-systems PhD who authored a maritime-trade forecasting model published via the US National Academy of Sciences. A researcher turned builder.
A pioneer of attention in speech, ex-Google Brain, and a co-author with deep-learning pioneer Geoffrey Hinton.
Theoretical computer scientist with 100+ papers - a PhD at 20, tenured at 23. The kind of resume that reads like a typo.
Pathway's customer list is short, specific, and unusually demanding - which is the point. These are organizations where being a second late is not a UX complaint but an operational failure.
NATO uses it to fold structured and unstructured data into adaptive situational awareness. La Poste leaned on it to optimize logistics, with reported total-cost-of-ownership cuts of up to 50%. Formula 1 teams use it to adapt strategy mid-race. Behind them: a developer community in more than 100 countries.
With Kadmos, Inovo.vc, Market One Capital, Id4, and angel Lukasz Kaiser - co-inventor of the Transformer.
From Inovo Venture Partners and Market One Capital, backing the original complex-systems thesis.
The timing is its own commentary. As Cohere and Writer mined the enterprise-AI arena, Pathway joined with a $10M seed and a sharper claim: the bottleneck isn't a bigger model, it's a model that can't remember. Investors who helped invent the current era are now funding the people trying to end its main limitation.
Estimated revenue sits around $3.4M - small, but the open-source flywheel (62,000+ stars) is the real moat the funding is meant to widen.
Return to the pit lane. The telemetry is still moving faster than any human can read it - but now something reads it back. The strategist isn't waiting for the lap to end; the decision and the data arrive together. At the postal depot, the next ten thousand parcels route themselves around a problem that hasn't fully formed yet. In the quiet room, the analyst's map updates while the situation is still a situation.
None of these people will say the word “framework.” That's the tell. Pathway's ambition isn't to be noticed - it's to close the gap between what happened and what you know until the gap stops mattering. The Baby Dragon Hatchling may or may not grow into the post-transformer future it's named for. But the smaller, stubborner idea underneath it - that AI should keep learning after the lesson ends - is already running, in more than a hundred countries, one live second at a time.