rePUSHTI is a member of NVIDIA Inception, NVIDIA's programme for companies building with AI. It gives us access to the compute, the tooling and the engineering support behind the agents that run inside the Travel OS.
What we are building with it
More than 70% of the booking re-verifications we run are handled agentically, end to end. Getting there took models that understand how a hotel actually answers the phone — in the local language, with a front-desk clerk who has thirty seconds and a queue behind you.
- Language models trained on the shape of real hotel and supplier conversations
- Agents that decide which channel a property will actually respond on
- Speech models for the voice calls that make up most re-verifications
- Evaluation harnesses, so an agent's answer is checked before a person ever relies on it
Why it is worth doing properly
A general-purpose model will happily tell a traveller that their booking is fine. The whole product depends on it never doing that unless a human at the property actually said so. That distinction is a training and evaluation problem, and it is where most of this work goes.
The agents are allowed to be fast. They are not allowed to be confident about something nobody confirmed.
