New opportunity

Principal Applied Scientist

Microsoft · Bangalore, India

About this role

Microsoft Advertising is building the next generation of AI systems for understanding advertiser behavior, detecting anomalies and emerging threats. We are looking for a Principal Applied Scientist with a strong foundation in mathematics, statistics, and core machine learning to advance foundation models, anomaly detection, decision uncertainty modeling, tool-using agents, and rigorous evaluation.

Responsibilities

Define and lead scientific initiatives in one or more areas eg foundation models, behavioral modeling, anomaly detection, threat modeling, agentic systems.
Develop scalable learning systems that understand entities, content, relationships, and behavior over time while identifying known, emerging, and previously unseen risks.
Develop methods to model and propagate uncertainty across individual models, model cascades, agent trajectories, retrieved evidence, automated decisions, and human judgments.
Use uncertainty, confidence, severity, and business impact to determine when to automate, gather additional evidence, invoke a more capable system, abstain, or escalate to expert review.
Translate threat models and adversarial insights into data strategies, learning objectives, model architectures, agent capabilities, and evaluation plans.
Advance the training, post-training, and evaluation of agents that use tools and evidence to investigate complex cases and produce grounded outcomes.
Address challenging learning settings involving distribution shift, sparse or delayed labels, noisy supervision, class imbalance, selective observation, and adaptive adversaries.
Translate scientific advances into reliable, efficient, and measurable production capabilities across Microsoft Advertising.
Provide technical leadership, mentor scientists, and influence the long-term architecture of AI-driven trust and safety systems.

Required qualifications

Bachelor’s, Master’s, or Doctorate degree in Computer Science, Mathematics, Statistics, Electrical Engineering, Operations Research, or a related quantitative field, with relevant industry or research experience .
Strong foundation in probability, statistics, linear algebra, optimization, numerical methods, experimental design, and statistical decision theory.
Deep expertise in modern machine learning, including foundation or representation learning, behavioral and temporal modeling, anomaly detection.
Proven experience in post-training and evaluating large-scale models.
Experience modeling uncertainty in production decision systems.
Ability to model threat and abuse scenarios.
Strong programming skills in Python and experience with frameworks such as PyTorch, JAX, TensorFlow, or equivalent technologies.
Proven ability to take scientific ideas from formulation through experimentation, production deployment, and measurable impact.
Demonstrated technical leadership through scientific direction, architecture, mentorship, and influence across science, engineering, product, and security teams.

Preferred qualifications

Experience with tool-using agents, retrieval, agent post-training, reward modeling, or trajectory evaluation.
Experience in trust and safety, fraud, abuse, cybersecurity, moderation, account integrity, or policy enforcement.
Experience working with temporal, multimodal, heterogeneous, or graph-structured data.
Strong publication or production track record in machine learning, agents, anomaly detection, probabilistic modeling, adversarial ML, multimodal learning, or trust and safety.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Machine LearningFoundation ModelsAnomaly DetectionThreat ModelingPythonPyTorchJAXTensorFlowGenerative AIAI AgentsProbabilistic ModelingAdversarial Machine Learning

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Advancing Foundation Models and Behavioral Understanding
Developing and scaling foundation models tailored for behavioral, content, entity, and risk understanding across large-scale multimodal, temporal, and relational data streams within Microsoft Advertising.
Decision Uncertainty Modeling and Automated Escalation
Building methods to propagate and leverage uncertainty across models, agent trajectories, retrieved evidence, and automated decisions to determine when to automate, gather more evidence, or escalate to human review.
Designing Tool-Using Agents for Threat Investigation
Advancing the training, post-training, and evaluation of tool-using agents that investigate complex abuse cases, gather evidence, and produce grounded, reliable outcomes against evolving adversaries.
Rigorous Evaluation and Productionization
Taking scientific ideas from formulation through experimentation, large-scale training, and production deployment while ensuring rigorous evaluation under distribution shift, class imbalance, and adversarial conditions.

Likely challenges to think through

Balancing Automation Confidence with High-Stakes Trust and Safety Risks
The challenge lies in reliably modeling decision uncertainty under sparse or delayed labels and adaptive adversaries, where false automation can lead to severe business or user impact. Preparation involves reviewing statistical decision theory, calibration methods, and robust cost-sensitive evaluation frameworks.
Generalizing Across Changing Adversarial Environments and Sparse Labels
Detecting unknown, emerging, and previously unseen abuse patterns amidst class imbalance and distribution shift. Preparation involves exploring anomaly detection strategies, semi-supervised learning techniques, and handling noisy supervision in adversarial settings.

Practice your answers

Technical practice

Modeling Decision Uncertainty in Production Systems

What to coverHow would you design a method to propagate uncertainty across model cascades and agent trajectories to determine when to automate a decision versus escalating to human review? Discuss how you factor in confidence scores, severity, and business impact.

Anomaly Detection Under Extreme Class Imbalance

What to coverWhat strategies do you use for detecting previously unseen abuse patterns and anomalies when dealing with extreme class imbalance and sparse or delayed labels? Explain how you handle distribution shift and adaptive adversaries.

Post-Training and Evaluation of Large-Scale Foundation Models

What to coverDescribe your approach to post-training and rigorous evaluation for large-scale models in behavioral or risk understanding tasks. Address how you evaluate model robustness, generalization across markets, and potential failure modes.

Tool-Using Agents for Complex Investigation

What to coverHow do you approach the training, post-training, and evaluation of tool-using agents that gather evidence and investigate complex cases? Discuss reward modeling, trajectory evaluation, and ensuring grounded outcomes.

Statistical Decision Theory and Optimization

What to coverHow do probability, statistics, and statistical decision theory inform your design of learning objectives and model architectures? Provide an example of how optimization and numerical methods apply to risk and threat modeling.

Handling Noisy Supervision and Selective Observation

What to coverWhen working with behavioral and temporal data subject to noisy supervision and selective observation, what learning paradigms or loss formulations do you employ to prevent overfitting to biased feedback loops?

Adversarial Machine Learning and Threat Modeling

What to coverHow do you translate threat models of emerging abuse scenarios into concrete data strategies and robust model architectures? Discuss techniques to defend against adaptive adversaries trying to bypass automated detection.

Large-Scale Machine Learning Architecture in Python

What to coverWhen implementing scalable learning systems in Python using frameworks like PyTorch or JAX, how do you architect training pipelines for high-throughput, multimodal, or relational data to ensure efficiency and production reliability?

Behavioural practice

Leading Complex Scientific Initiatives from Formulation to Production

What to coverDescribe a situation where you took a high-ambiguity scientific idea from initial problem formulation through experimentation and production deployment. Discuss how you measured product impact and handled technical hurdles along the way.

Cross-Functional Technical Leadership and Mentorship

What to coverGive an example of how you provided technical leadership and mentored scientists while influencing architecture across science, engineering, and security teams. How did you align conflicting priorities?

Navigating High-Stakes Trade-Offs in Trust and Safety

What to coverDescribe a time when you had to balance automated decision speed and business impact against the risk of false positives or abuse in a sensitive domain. How did you determine the threshold for human escalation?

Translating Adversarial Insights into Scalable Learning Systems

What to coverShare an instance where you identified an emerging threat or novel abuse pattern and translated those insights into concrete model architecture and evaluation plans. What was the outcome?

Sharpen your knowledge

Topics to prepare

Decision Uncertainty Modeling and Propagation
Review mathematical frameworks for modeling and propagating uncertainty across individual models, model cascades, agent trajectories, and human judgment. Exercise: Design a probabilistic decision pipeline that combines model confidence scores with cost matrices to dynamically automate decisions or escalate to expert review. Relevance: Directly addresses the core responsibility of building reliable trust and safety decision systems in Microsoft Advertising.
Anomaly Detection and Adversarial Threat Modeling
Study techniques for detecting novel, evolving abuse patterns, distribution shift, and sparse labels under adversarial settings. Exercise: Formulate an anomaly detection benchmark on temporal/behavioral data with injected class imbalance and test semi-supervised learning methods. Relevance: Matches requirements for identifying emerging risks and malicious advertiser behavior.
Tool-Using Agents and Post-Training Evaluation
Examine agent post-training, reward modeling, retrieval integration, and trajectory evaluation. Exercise: Set up an evaluation harness to measure grounded outcomes and tool-use reliability for an agent investigating multi-step interaction traces. Relevance: Directly aligns with preferred and required qualifications for agentic system design.

Skills to revise

PyTorch and Large-Scale Model Post-Training
Practice implementing post-training, representation learning, and fine-tuning pipelines using PyTorch or JAX. Exercise: Fine-tune a transformer-based encoder on sequential user/advertiser interaction logs using custom loss functions designed to handle noisy supervision. Relevance: Essential for hands-on development of foundation models and scalable learning systems using required programming stacks.

Your preparation checklist

Prepare a End-to-End ML Project Walkthrough
Assemble a detailed technical presentation of a past project demonstrating end-to-end ownership from problem formulation through large-scale training, deployment, and measurable impact, specifically highlighting foundational models, anomaly detection, or decision systems.
Draft Uncertainty and Risk Modeling Frameworks
Prepare case studies or architectural notes illustrating how you have modeled decision uncertainty, propagated confidence scores across model cascades or agent trajectories, and designed automated versus human-in-the-loop escalation criteria.
Build a Tool-Using Agent Evaluation Portfolio
Document your experience with agentic systems, post-training, reward modeling, and trajectory evaluation, outlining how you design evaluation harnesses for tool-using agents investigating complex behavioral cases.
Review Mathematical and Statistical Foundations
Refresh core concepts in probability, statistics, linear algebra, optimization, numerical methods, and statistical decision theory to prepare for technical discussions on experimental design and adversarial learning settings.
Prepare Evidence of Technical Leadership and Mentorship
Compile examples of how you have provided scientific direction, mentored scientists, and influenced cross-functional architecture and long-term strategy across science, engineering, product, and security teams.
Formulate Thoughtful Employer Questions
Develop specific questions about Microsoft Advertising's architecture for handling adaptive adversaries, scaling tool-using agents, and balancing automated decisions with human review in trust and safety systems.

Availability has not yet been verified. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in India. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…