• Teams Subtitles
  • Teams Subtitles provides real-time simultaneous speech translations for 40 spoken languages, into 34 target languages. This constitutes access to real-time translations for the 270M+ MaUs of Teams and represents one of the greatest language barrier removers in the technology industry in terms of its coverage (users and languages). I led the project since its beginning, from training and designing the first models to conducting the final human evaluations prior to shipping. Novel algorithms designed and implemented by me in this project led to large gains in the final user experience and are directly consumed in the shipped feature. Some of the public-facing aspects of this work are under preparation for publications. Meanwhile, this feature release within Teams was announced at Microsoft Ignite’22 and widely covered in the technology media as well, e.g., Computer World, TechWire, IT World, Akari.
• Microsoft Translator
  • Microsoft Translator (belonging to the wider suite of Azure Cognitive Services) is the translator API that powers the consumer-oriented Bing Translator and the enterprise-oriented Azure Cognitive Services Translator API. My work has contributed to reducing hallucinations 10x across different language pairs, in the reduction of certain long tail errors by 4x and in improving the reliability of the underlying neural translation models in multiple ways. The work on hallucinations was recognized as one of the notable research works across Microsoft by MSR and the Office of the CTO in 2021. Some of the public-facing aspects of this work are introduced in my blog post.
• Document Translator
  • The Document translator is powered by the industry-first deployment of Mixture of Experts (MoE) models for MT. I led the long-tailed error measurements and data-augmentation for robustness aspects of the project across different language families, among other things. This was widely covered in the news as well, e.g., TechCrunch, VentureBeat, ZDNet, MSR Blog Post.
• Dynamics-365
  • I designed a key piece of technology that is used to verify the fidelity of responses from Large Language Models in Microsoft Dynamics 365 suite of products, targeted towards enterprise marketing scenarios.
• Telemetry++
  • A Microsoft internal-only product/utility, this allowed engineers to leverage telemetry data for data-driven decision making. The work was presented at the Grace Hopper Conference in India and also won recognition as the best Hackathon project in the org.