Aug 2025 – Present · 9 mos · Vancouver, British Columbia, Canada

• Did you know, that OneDrive and SharePoint,
• ✅ Has over 1.25 billion active users worldwide
• ✅ Stores at least 1.2 million terabytes of data (that’s about 1.2 exabytes of files)
• ✅Orgs add over 2 billion pieces of content every day
• With so many files uploaded, shared, edited, and synced every second, the question is: how do we ensure that amidst all this content, the right content is surfaced, at the right time, to the right user?

Sep 2023 – Aug 2025 · 1 yr 11 mos · Vancouver, British Columbia, Canada

• AgentSOC Research to Production: Built reusable, scalable post-training pipelines across multi-tenant environments, accelerating iteration for model customization and optimization. Our RL based finetuned model improved the alert triaging precision by 4x. As a part of this end-to-end pipeline, collaborated with multiple teams to curate high-quality labelled data across multiple scenarios for evaluation and continuous benchmarking.
• AI Research for security:
• ✅ built a multi-agent Autogen system that improved PyRIT red-teaming success by 40% (https://github.com/Azure/PyRIT)
• ✅ drove research with an IJCAI 2025 submission on explainable jailbreak defenses (https://arxiv.org/abs/2504.12321)
• ✅ two accepted MLADS 2025 papers, an MSJAR paper,
• ✅ open-sourced DefenderBench toolkit (https://github.com/microsoft/DefenderBench);
• ✅ along with 4 Microsoft patents in the space of security and AI

Sep 2021 – Aug 2023 · 1 yr 11 mos · Vancouver, British Columbia, Canada

• Using state-of-art deep learning techniques to improve the security of Microsoft products.
• ✅ Built a scalable, compliant ML training pipeline; processing ~10M+ messages/day with automated cleaning and fully reproducible runs—creating a reusable foundation for training ML models for several applications.
• ✅ Experimented with an eyes-off, LLM-based phishing detector using unstructured free text content, delivering 41% improvement of the previously production system
• ✅ Productionized the LLM via model optimization and quantization for remote inference on live traffic, reliably serving ~200 RPS per instance with monitoring and safety gates.