Apr 2023 – Present · 2 yrs 11 mos

• Tech Lead across several initiatives in Trust Org
• Data Retrieval Stack for LLMs & AI Agents
• Building a horizontal, unified stack for retrieving structured and unstructured data (SQL, offline logs, K-V stores, documents, etc.) for LLM powered AI agents.
• Standardizing and indexing data into LLM-friendly formats to support reasoning and summarization.
• Powering Retrieval-Augmented Generation (RAG) across all GenAI applications.
• Trust in GenAI Products
• Designing systems to detect and prevent content abuse in GenAI outputs, including:
• Gender and racial bias,
• Prompt injection and jailbreak attacks,
• Harmful or policy-violating generations.
• Feed Ranking Infrastructure
• Architected a highly available, near real-time platform to rank LinkedIn Feed using content and member signals.
• Supports 100K+ QPS writes and 4M+ QPS reads with low latency and high durability.
• Asynchronous Orchestration Framework
• Developed a distributed DAG-based framework to execute processing pipelines (e.g., OCR, feature extraction, classifiers) asynchronously.
• Enabled seamless onboarding of new processors with dependency management (e.g., run classifier after OCR and caption extraction).
• Scales across distributed nodes for high-throughput, fault-tolerant execution.

Oct 2020 – Mar 2023 · 2 yrs 5 mos

• Content Abuse ML Platform
• Designed a scalable, multi-tenant platform supporting 100+ lines of business to host text, image, video, and live-stream classification models.
• Optimized memory, throughput, and latency—cut serving cost by 13x vs. legacy system.
• Built a self-serve launch platform saving 50+ eng-days/year.
• Received Craftsmanship Award for LinkedIn’s image/video classification stack.
• Powers 2B+ inferences/week across 50+ production models.
• Gray Content Initiative
• Built platform support to detect and act on nuanced “gray” content, enabling policy-aware moderation beyond binary classification.
• Abuse War Rooms
• Prevented misinformation spread during the 2020 US Elections.
• Investigated and mitigated “fake romance” abuse campaigns targeting members.
• Leadership
• Led multiple teams across QTR initiatives.
• Created hiring criteria and drove interview processes for new Trust roles.
• Launched cross-team knowledge-sharing initiatives across Abuse and Privacy teams.

Aug 2018 – Sep 2020 · 2 yrs 1 mo

• Content Abuse (Safety AI Platform)
• Part of infrastructure initiatives to host and operationalize ML models for detecting abusive content across text, images, and videos.
• Video Classification at Scale: Engineered a pipeline to classify videos up to 6 hours long. Migrated workloads from AWS to LinkedIn’s internal infra—achieving a 3.6x cost reduction and saving ~$175K/year.
• Vision + Meta-feature Models: Integrated OCR and caption-based signals into image classifiers, significantly improving precision and recall in content abuse detection.
• Near-Dupe Detection: Deployed infrastructure to host image-near-duplicate models for high-throughput visual similarity filtering.
• Rapid Threshold Tuning: Cut ML iteration cycles from ~3 days to ~10 minutes through improvements to classifier threshold management.
• Codeless Model Integration: Built a declarative integration framework that reduced new model onboarding time from ~3 days to ~2 hours.
• Account Access (Data Privacy & Member Lifecycle)
• Designed scalable systems that manage member account lifecycle and ensure compliance with global privacy regulations.
• Selective Data Deletion: Architected a new feature enabling members to delete specific sensitive data while retaining their accounts.
• GDPR-Compliant Data Export: Modernized member data export workflows across LinkedIn products in alignment with evolving GDPR mandates.
• Account Hibernation: Built the infrastructure for temporary account deactivation (“hibernation”) to give members more control over their presence on the platform.
• Awards
• 🏆 Winner, LinkedIn AI Hackathon — for building a topic tag prediction system for image content