Two PhD Candidates in Datadog's Paris AI Lab Help Build Toto, a Timeseries Foundation Model

Datadog’s AI Research Lab in Paris expanded in 2024 through a partnership with France’s CIFRE program and recruited doctoral candidates to pursue production — focused research tied to real engineering problems. The CIFRE (Conventions Industrielles de Formation par la Recherche) arrangement is government — funded and supports three — year industry — academia doctoral projects, combining university supervision with day-to-day collaboration with engineering and product teams across borders. Two PhD researchers highlighted at the lab, Viktoriya Zhukova and Salahidine Lemaachi, work on Toto, an open-source timeseries foundation model that has been downloaded more than 9 million times. Their projects are academically rigorous but explicitly designed so outcomes can move into production pipelines and open-source observability tooling.

Viktoriya concentrates on multimodal timeseries forecasting, enriching numeric series with additional data modalities to improve prediction accuracy. She splits her time between the Paris office and Université Paris — Saclay, meets her academic supervisor weekly, and balances reading papers, running experiments, and preparing publications alongside engineering collaboration. Salahidine’s research focuses on building world models with an emphasis on real-world applicability rather than purely theoretical advances. His work aims to produce models and components that can be evaluated and iterated on using operational telemetry and concrete use cases from Datadog’s product and engineering teams.

The lab stresses access to large, diverse observability datasets and a roughly 20‑person research group spanning Paris and New York, giving it a scale advantage over many purely academic teams. That access lets researchers test methods against production telemetry early, so deployment considerations and practical constraints inform model design from the start. For engineers and builders, that pipeline matters: models developed in the lab are evaluated against live telemetry and real use cases, increasing the likelihood that research outputs will include deployment — ready components. Viktoriya points to practical forecasting examples — such as estimating daily GPU consumption — where multimodality can materially improve resource planning and alerting.

Researchers say the Paris tech ecosystem and the CIFRE framework shape their experience by combining industry exposure, academic rigor, and cross — disciplinary collaboration. The lab’s intention is to accelerate transitions from papers to product, producing tooling and models that practitioners can adopt or extend within observability stacks and open-source projects like Toto.