About the Role
<div>
<h2>About Scality</h2>
<div>Scality builds data infrastructure software for enterprise AI, cyber resilience, and sovereign control at multi-petabyte to exabyte scale. Its AI Data Infrastructure platform reduces operational burden through AI-powered autonomous operations while aligning the right storage media, performance, and protection to each workload through human-approved lifecycle policies.</div>
<div> </div>
<div>Built on CORE5 cyber resilience and open-code principles, Scality software delivers extreme performance, operational simplicity, and sustainable economics. </div>
<div> </div>
<div>The world's most demanding enterprises and government organizations depend on Scality to power AI initiatives, defend critical data, and build infrastructure that lasts for decades. Recognized as a leader by Gartner, Scality is where AI remembers, learns, and thinks. F</div>
<div> </div>
<div>Follow us on LinkedIn. Visit <a href="https://www.scality.com/">www.scality.com</a> and our blog.</div>
</div><br><strong>Context & Motivation</strong><br><div>A distributed storage cluster continuously emits telemetry at every level, from the platform and S3 service down to each server and disk. A failing disk, server or storage node leaves traces in these multivariate time series, from subtle shifts in a few metrics to system-wide disruption. Detecting and diagnosing these anomalies early is critical for:</div>
<li>Predictive maintenance: isolating or replacing a component before it impacts service</li>
<li>Faster diagnosis: pointing support engineers to the root cause, not just to an alert</li>
<li>Higher availability and durability of customer data</li>
<div>This internship is part of a joint research project between Scality and an INRIA research team. The project released SHAD (Scality High-dimensional Anomaly Detection), an open benchmark published at NeurIPS 2026 (Evaluations & Datasets Track). It contains 215 real 171-dimensional time series from Scality RING clusters, annotated for injected disk, server and storage-node failures. SHAD's baselines show the problem is far from solved. No detector handles every failure type, per-dimension attributions find root causes no better than chance, and off-the-shelf LLMs cannot interpret the anomalies. This internship builds on SHAD to close these gaps.</div>
<div><br>More details on the project: <a href="https://scality.github.io/shad">scality.github.io/shad</a></div><br><strong>Your Mission</strong><br><div>As part of our R&D team and in close collaboration with INRIA researchers, you will develop AI models that detect, explain and interpret anomalies in storage-cluster telemetry. Your work will include: </div>
<div> </div>
<div>1. State of the Art</div>
<div>Review the literature and open-source projects on multivariate time series anomaly detection, root-cause attribution and LLM-based interpretation of time series. Also survey recent work on AI agents that automate research and experimentation. Starting points include SHAD and the TSB-AD benchmark. </div>
<div> </div>
<div>2. Anomaly Detection & Diagnosis Models</div>
<div>Design, train and benchmark models that improve on SHAD's baselines at each stage of the pipeline:</div>
<li>Detection that stays accurate across disk, server and storage-node failures (e.g., ensembles, hybrid models, time series foundation models)</li>
<li>Explanation: attributing each anomaly to the dimensions that actually cause it</li>
<li>Interpretation: turning detections into actionable diagnoses (failure type, criticality, affected component) </li>
<li>Early detection of progressive degradations, the core of predictive maintenance</li>
<div> </div>
<div>3. Fleets of AI Agents for Research</div>
<div>Set up and manage fleets of AI agents that run your studies at scale:</div>
<li>Design agent workflows that launch experiments, run benchmarks and compare methods</li>
<li>Orchestrate and monitor many agents working in parallel</li>
<li>Build safeguards for rigor: experiment tracking, reproducibility, independent verification of results</li>
<li>Assess what agent-driven research changes in practice: speed, coverage, failure modes</li>
<div> </div>
<div>4. Evaluation & Validation</div>
<div>Evaluate every model with rigorous, reproducible protocols. Use them to:</div>
<li>Measure each stage with SHAD's metrics (VUS-PR for detection, NDCG for explanation, a questionnaire for interpretation)<br>Quantify performance per failure family and scenario</li>
<li>Identify which approaches are ready to transfer to production monitoring</li>
<br><strong>Expected Outcomes & Valorization</strong><br><div>Depending on results, the work may be valorized through:</div>
<div> </div>
<li>A scientific publication with the Scality and INRIA researchers</li>
<li>New methods and baselines