Experience

Ads Trust & Safety

ATS is responsible for ensuring that ads served across Microsoft's platforms comply with policy — every violation that reaches a user is a defect. Keeping that rate low is the core mandate of the team, and measuring it accurately is what makes enforcement possible.

I own the defect rate measurement pipelines — the infrastructure that tracks what fraction of served ads are in policy violation, across both Search and Native platforms (MSN.com and partner network). These pipelines are the source of truth for ATS: the number they produce determines whether enforcement is working end-to-end.

I also rewrote the clickbait measurement model, updating it to align with latest policy guidelines. Offline: first-party precision improved from 43.3 → 73.5, third-party F1 from 72.7 → 76.4. Shadow runs confirmed the gains: human reviewers sided 57% vs 43% with the new system. I also built a structured knowledge base so an LLM agent can now execute 70–80% of the analysis pipeline end-to-end.

Microsoft Audience Network (MSAN)

MSAN is Microsoft's programmatic ad network, serving ads across Bing, MSN.com, and partner sites. Retrieval is the first — and most consequential — stage of the ad pipeline: it determines which ads are even in contention for a given query.

I worked on intent-based retrieval: building encoder models that understand what a user is actually looking for and surface ads that genuinely match. The next-generation model I led achieved +14.32 absolute precision points (P@100) over the production baseline using 30× less training data, delivering +0.61% CTR and +0.51% revenue lift in A/B testing. The evaluation methodology I built was adopted as standard across US and India MSAN teams.

I also designed a unified evaluation framework across three teams (IDC, STCA, US MSAN) and built a high-impact product ads index using ML-driven retrieval that replaced the legacy rule-based system, delivering a 2.77% revenue lift.

Built an image quality scorer — a fine-tuned ResNet-50 (F1: 90%) that now evaluates nearly every food photo on the platform. Also designed an automated ad creation system using generative models, and built Photo Cake — a real-time image overlay feature that drove 3,000+ orders on Mother's Day.

Led semantic text similarity at scale — scaled to 1M+ texts with F1 of 85%, exploring prompt engineering, fine-tuning, and novel loss functions. Wrote 20+ pages of data preparation guidelines and supervised the annotation team.