New opportunity

Senior Applied Scientist - Ads Ranking & Retrieval

Microsoft · Bangalore, India

About this role

Microsoft Ads seeks a Senior Applied Scientist to advance ad retrieval, matching, ranking, and generation at web scale. The role combines research with production delivery, aiming to improve user experience, advertiser return on investment, and platform efficiency. It is based in Bangalore, Karnataka, India, with four days per week in the office and less than 25% travel.

Responsibilities

Research and develop retrieval, ranking, matching, and generative models. Train, fine-tune, align, improve, and productionize small language models, large language models, and large reasoning models.
Improve the Ads ranking platform’s usability, reliability, scalability, efficiency, and architectural coherence. Ship solutions, measure their impact, and iterate on real-world outcomes.
Set technical direction, coach a distributed team, influence cross-organization strategy, follow AI research trends, and collaborate with research and engineering teams.

Required qualifications

A bachelor’s degree in Statistics, Econometrics, Computer Science, Electrical or Computer Engineering, or a related field with 4+ years of related experience; a master’s degree in one of these fields with 3+ years; a doctorate with 1+ year; or equivalent experience. Related experience includes statistics, predictive analytics, or research.
Ability to meet applicable Microsoft, customer, or government security-screening requirements. The Microsoft Cloud background check is required upon hire or transfer and every two years thereafter.

Preferred qualifications

A relevant master’s degree with 5+ years of related experience, a doctorate with 3+ years, or equivalent experience. The posting separately cites 6+ and 5+ years of machine-learning experience shipping large-scale models to production.
Expertise in training and inference optimization; experience with SLM, LLM, and LRM training, fine-tuning, and post-training; and experience scaling recommendation systems with massive query and item spaces and multi-stage ranking pipelines.
Proficiency with PyTorch, Hugging Face, or TensorFlow and distributed training on large datasets; ability to influence platform architecture and cross-team roadmaps; and publications in venues such as NeurIPS, ICML, KDD, WWW, ACL, or SIGIR.

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

Skills for this role

Machine LearningAd RetrievalAd RankingRecommendation SystemsMatching ModelsGenerative ModelsLarge Language ModelsSmall Language ModelsLarge Reasoning ModelsModel TrainingFine-TuningModel AlignmentPost-TrainingTraining OptimizationInference OptimizationMulti-Stage RankingDistributed TrainingPyTorchHugging FaceTensorFlowTechnical LeadershipResearch Collaboration

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

Advance ad retrieval and ranking
The role calls for research and development across ad retrieval, matching, ranking, and generative models. Preferred experience includes recommendation systems with massive query and item spaces and multi-stage ranking pipelines.
Bring language models into production
The posting asks for training, fine-tuning, aligning, improving, and productionizing SLMs, LLMs, and LRMs, with preferred experience in post-training and training and inference optimization.
Deliver measurable product impact
The scientist is expected to ship solutions, measure their impact, and iterate on real-world outcomes that improve user experience, advertiser ROI, and platform efficiency.
Improve the ranking platform
A stated responsibility is to improve the platform’s usability, reliability, scalability, efficiency, and architectural coherence—not only model quality.
Set technical direction across teams
This senior individual-contributor role includes setting project direction, coaching a distributed team, collaborating with research and engineering, and influencing cross-organization strategy.

Likely challenges to think through

Quality versus serving cost
Likely preparation challenge, not a confirmed employer problem: more capable models may improve retrieval or ranking while increasing inference cost or latency at the stated web scale. Prepare to explain how you would compare model quality with serving efficiency and decide where a model belongs in a multi-stage pipeline.
Balancing competing outcomes
Likely preparation challenge: a change that helps advertiser ROI may not improve user experience or platform efficiency equally. Prepare a concrete approach to measuring all three outcomes and deciding when a gain in one does not justify a loss in another.
Research speed versus platform coherence
Likely preparation challenge: shipping a novel approach quickly can conflict with reliability and a coherent architecture, particularly when research and engineering teams must align. Prepare an example of how you would evaluate a promising idea, identify production risks, and build agreement on its rollout or revision.

Practice your answers

Technical practice

How would you design ad retrieval for a massive query and item space?

What to coverSuggested answer focus: Explain how you would generate relevant candidates at scale, then assess retrieval quality, coverage, and computational cost. Discuss how retrieval choices affect the later matching and ranking stages.

How would you improve the quality of query-to-ad matching?

What to coverSuggested answer focus: Describe how you would define a useful match and compare modeling approaches using evidence from real-world outcomes. Address the trade-off between broader matching and showing less relevant ads.

How would you structure and evaluate a multi-stage ad ranking pipeline?

What to coverSuggested answer focus: Explain what each stage should accomplish and how you would allocate complexity across stages. Discuss how you would evaluate user experience, advertiser ROI, and platform efficiency together.

Where could generative models add value in ad systems?

What to coverSuggested answer focus: Identify a plausible application to ad generation or matching, and explain how you would test its value against a simpler approach. Cover quality, serving cost, and safeguards against unsuitable outputs as proposed evaluation considerations, not stated platform requirements.

How would you choose among SLMs, LLMs, and LRMs for a production use case?

What to coverSuggested answer focus: Compare task quality with training and inference demands, and explain when fine-tuning or alignment might be worthwhile. Describe how you would verify that model improvements persist in production.

How would you approach distributed training on a large dataset?

What to coverSuggested answer focus: Discuss data handling, training efficiency, and reproducibility using a relevant deep-learning framework such as PyTorch or TensorFlow. Explain how you would diagnose a scaling bottleneck before changing the training setup.

How would you optimize model inference without losing too much quality?

What to coverSuggested answer focus: Identify where serving cost or response time arises and propose changes to measure against a quality baseline. Make the quality–efficiency trade-off explicit for a system serving requests at web scale.

How would you troubleshoot reliability or scalability problems in a ranking platform?

What to coverSuggested answer focus: Explain how you would isolate the problem across model behavior and platform operation before proposing a fix. Discuss how to check that the fix improves reliability without undermining usability or architectural coherence.

How would you measure whether a shipped model improved the Ads product?

What to coverSuggested answer focus: Define success in terms of user experience, advertiser ROI, and platform efficiency, as named in the posting. Explain how you would interpret conflicting results and decide what to iterate on next.

How would you decide whether new AI research is ready for production?

What to coverSuggested answer focus: Describe how you would compare a research idea with existing approaches on model quality and practical deployment costs. Explain what evidence would justify continued research, a limited production trial, or stopping the work.

Behavioural practice

Tell me about a time you set technical direction for an ambiguous project.

What to coverSuggested answer focus: Use a real situation, your decision-making actions, and the resulting outcome. Explain how you balanced scientific opportunity with a deliverable that could be measured in practice.

Describe a time you coached colleagues working across locations.

What to coverSuggested answer focus: Give a real situation, the specific coaching or coordination actions you took, and the result. Distinguish your contribution from the team’s work without claiming sole credit.

Tell me about a time you influenced an architecture or roadmap decision across teams.

What to coverSuggested answer focus: Describe the competing priorities, the evidence and actions you used to build alignment, and what changed. Include a real result, even if the proposal was only partly adopted.

Describe a collaboration between research and engineering that changed your approach.

What to coverSuggested answer focus: Explain the situation, how you worked through differing technical perspectives, and the outcome. Highlight a concrete adjustment you made rather than presenting collaboration as agreement alone.

Tell me about a solution you shipped and then revised based on its results.

What to coverSuggested answer focus: State the original goal, your actions to measure real-world impact, and what the results showed. Explain the subsequent iteration and its outcome, including any trade-offs that remained.

Sharpen your knowledge

Topics to prepare

Retrieval, matching, and multi-stage ranking
The posting explicitly covers ad retrieval, matching, ranking, and recommendation systems with massive query and item spaces. Suggested revision exercise: sketch a candidate-retrieval-to-ranking pipeline and explain what each stage optimizes and where relevant ads might be lost.
SLM, LLM, and LRM adaptation
The role explicitly calls for training, fine-tuning, aligning, post-training, and productionizing these models. Suggested revision exercise: compare when each model type might be useful for ad matching or generation, and outline how you would evaluate an adapted model before deployment.
Training and inference efficiency
Training and inference optimization are preferred qualifications, and the platform must scale efficiently. Suggested revision exercise: work through a model-serving bottleneck and explain the trade-offs among model quality, latency, and compute cost.
Measuring product impact
The overview emphasizes shipping, measuring impact, and iterating to improve user experience, advertiser ROI, and platform efficiency. Suggested revision exercise: propose an evaluation plan for a ranking change, separating model-quality measures from real-world outcomes and potential trade-offs.
Platform architecture and scientific leadership
The role calls for architectural coherence, technical direction, coaching, and cross-organization influence. Suggested revision exercise: prepare a design-review outline that states a research hypothesis, production risks, architectural trade-offs, and how research and engineering teams would coordinate.

Skills to revise

Prototype a multi-stage recommender
Suggested hands-on exercise: build a small illustrative candidate-retrieval and ranking flow over sample query–item data, then inspect candidates missed at each stage. This practices the posting’s explicit multi-stage ranking and large query/item-space work; the sample setup is preparation, not a stated employer requirement.
Fine-tune and evaluate a model
Suggested hands-on exercise: use one framework named in the posting—PyTorch, Hugging Face, or TensorFlow—to fine-tune a small model on an illustrative matching task and compare it with an unfine-tuned baseline. The role explicitly mentions these frameworks and SLM/LLM/LRM fine-tuning; this exercise does not imply a prescribed model or dataset.
Diagnose training at scale
Suggested hands-on exercise: profile a training run on a larger sample dataset and document bottlenecks, scaling trade-offs, and a distributed-training plan. This maps to the preferred experience with large datasets, distributed training, and training optimization.
Test inference improvements
Suggested hands-on exercise: benchmark an illustrative model before and after one inference optimization, recording quality, latency, and resource-use changes. Inference optimization and platform efficiency are explicit priorities; the benchmark design is a preparation suggestion.
Present a production decision
Suggested hands-on exercise: write a brief technical proposal for a retrieval or ranking change, including an impact-measurement plan, reliability and scalability risks, and decisions needing cross-team agreement. This reflects the posting’s emphasis on production delivery, platform architecture, and technical leadership.

Your preparation checklist

Build an ads-ranking project walkthrough
The posting emphasizes retrieval, matching, ranking, and production delivery. Prepare one relevant example, if available, showing the problem, model choices, deployment decisions, measured results, and what changed after launch.
Practice a multi-stage system design
As a preparation exercise, sketch candidate retrieval through final ranking for a massive query and item space. Explain trade-offs in relevance, latency, scalability, reliability, and platform efficiency; the posting does not specify an interview format.
Prepare a language-model technical deep dive
The role covers SLMs, LLMs, and LRMs, including training, fine-tuning, alignment, and productionization. Prepare an example you can discuss in depth, or a clearly labeled design proposal, with training and inference trade-offs.
Document how you measure impact
The posting calls for shipping, measuring, and iterating to improve user experience, advertiser ROI, and platform efficiency. Prepare a concise account of metrics, evaluation decisions, observed outcomes, and limitations from a relevant project, if available.
Gather evidence of research and engineering depth
Organize applicable examples of distributed training, large datasets, PyTorch, Hugging Face or TensorFlow, and training or inference optimization. Have publications ready to discuss if relevant; publications are listed as preferred, not required.
Rehearse a technical-leadership example
The role includes setting direction, coaching a distributed team, and influencing cross-organization strategy. Prepare one specific account of a technical decision, how others were involved, and the outcome, if you have such an example.
Write questions for the employer
Ask which retrieval or ranking problems are the current priority, how research ideas reach production, and how impact and platform trade-offs are assessed. You can also clarify the stated four-day in-office schedule and security-screening process.

Availability last checked: 17 Oct 2026. 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…