Changepoint detection is one of the grand challenges of machine learning [Truong et al. 2018,
Wang & Xie 2022]. As technology develops, the demand for changepoint procedures capable of
processing massive datasets and monitoring complex systems in real-time (online) has increased (e.g.,
trades and order-flows in finance, health data from smartwatches...).

The challenge is to detect changes in the distribution of the data over time. The archetipical
examples are changes in the mean, but there are more intricate applications: e.g., changes in the variance,
network structure, or images. State-of-the-art approaches optimise a measure of fit to the data. Over the
past 70 years, online & offline (processing data once it is complete) approaches have evolved separately
[Truong et al. 2018, Wang & Xie 2022].

We discovered a link between changepoint detection and geometry, bridging the gap between online
& offline approaches [Pishchagina et al. 2025]. For low-dimensional models in the exponential family
with up to 6 parameters, this link revolutionizes online maximum likelihood inference, matching the
speed of standard local optimisation methods for detecting a single change. Until recently, maximum
likelihood approaches were relegated to offline applications.

Further exploring this link, we aim to bring the power of offline approaches to online applications
and scale up changepoint inference for seamless analysis of more than a million for low and high-
dimensionnal models with latent variables for one or multiple changepoints.