The Behavior AI team builds the AI-based anomaly detection behind Datadog's security products. Our models learn what normal looks like across the billions of logs, events, and telemetry records flowing through the platform every second, and they flag the behavior that does not fit, on every record, in real time, at a cost that makes sense at our scale. What we build does not ship to a single feature. The same models power detection across many of Datadog's security products at once, so the work has impact well beyond any one team. Large general-purpose models are too slow and too expensive to run in that path, so we take the opposite approach: small, custom models, designed for high-throughput stream processing and optimized to run cheaply on every record.
We are hiring a Senior Applied Scientist to build these models from start to finish. You will contribute to designing the architecture, training at scale, and the optimization work that takes a model from training to running efficiently on production traffic. This optimization requires a deep understanding of the constraints imposed by both the software and the hardware, together with the applied mathematics to work within them: often it comes down to finding a mathematical reformulation that fits those constraints better, and that judgment can decide whether a model reaches production at all.
The work involves a number of open questions. How do you obtain most of the quality of a large model from one that is far smaller and cheap enough to run on the full stream? Where is it worth trading exactness for speed, and how do you reason about the error you accept? How do you make a small model's outputs clear enough that the detection engineers and analysts who rely on it can trust what it reports? If these are the problems you want to work on, we would like to hear from you.
At Datadog, we place value in our office culture: the relationships and collaboration it builds, and the creativity it brings to the table. We operate as a hybrid workplace to ensure our Datadogs can create a work-life harmony that best fits them.
What You'll Do:
- Design and build custom mid-size models for high-throughput stream processing, and train them at scale.
- Optimize these models from start to finish, working across both the mathematics and the systems, often by finding mathematical reformulations that fit the hardware and software constraints better.
- Build the training data pipelines the work depends on when they do not already exist, from the raw stream to a training-ready dataset.
- Work with engineering to integrate models into production, with a clear focus on GPU utilization, latency, and cost per record on live traffic.
- Plan the roadmap of model and system improvements, based on a solid understanding of the product and of what matters most to users.
- Build an agentic layer on top of the models to analyze, validate, and act on their outputs.
- Build lightweight interpretability tools that make model behavior easier to explain to the people who rely on it.
- Maintain and monitor the models, services, and infrastructure your team owns, and take part in your team's on-call rotation.
Who You Are:
- You have a BS/MS/PhD in Computer Science, Engineering, Machine Learning, Applied Mathematics, or a related scientific field, or equivalent experience.
- You have hands-on experience training and fine-tuning models at scale, and deploying them into production systems with real throughput and cost constraints.
More roles at Datadog
See Datadog & thousands more on the map
Every tech & IT company hiring across India — with AI match scores — on one live map.
Open the map →