Reshmi Ghosh


Ph.D. Applied Scientist, Microsoft (MSAI)


I’m Reshmi, Applied Scientist at Microsoft’s Search and Intelligent Assistance (MSAI) team, focusing on developing state-of-the-art Responsible AI and Trustworthy evaluation mechanisms for LLM applications. I have focused my efforts in the past year on identifying, evaluating, and mitigating security vulnerabilities (prompt injections) in M365 Chat CoPilot, extending to Office 365 CoPilots (PowerPoint, Excel, Word, Outlook, Teams, etc.). Shortly before that, I was also responsible for leading the efforts to architect and integrate Generative AI capabilities in Microsoft’s Productivity applications (M365 CoPilots, see Satya Nadella’s announcement of these products here (patent filed).

Before joining MSAI, I was selected to be part of the highly coveted AI rotational program at Microsoft’s New England Research and Development Center (NERD) (currently under the Office of CTO, previously under Jaime Teevan’s Office of Applied Research). I was one of the 17 individuals from 10K applicants (~0.2% acceptance rate) to this esteemed program. During my time in MAIDAP, I spent time bringing ML capabilities to detect anomalies in Azure services, serviced intelligent commanding features to Microsoft Office using seq2seq language modeling as well as researching user preferences in UX positioning by developing a novel loss function. I furthermore worked with the Viva Topics team to develop new probabilistic graphical methods to measure time saved to find new knowledge and developed an LLM application and novel evaluation framework for generating rich-text formatting within a client application.

Prior to joining Microsoft, I graduated with a Ph.D. from Carnegie Mellon University in 2021. I have always been passionate about machine learning, deep learning, economics, climate change, and renewable energy. I have 5+ years of working with data-driven techniques in python to discover underlying patterns in large datasets. I love leveraging data science techniques to derive key actionable insights about business and societal problems, and translating these insights into business strategies.

My doctoral research reflects the blend of of my interest areas in a public policy setting for energy systems. I am analyzing the future reliability of power systems for the United States by developing probabilistic mathematical models using terrabytes of reanalysis datasets to measure the resource adequacy metrics. Additionally, I am also working to reconstruct decades of missing electricity demand data using large seq2seq models for detecting anomalous behavior (vanilla and with Attention). I am also attempting to use causal inference methods to identify underlying patterns in multi-decadal electricity load data, and understand it’s implication on grid system.

Ph.D. Dissertation Resume LinkedIn Github Twitter



[June 2024] Paper from IEEE Secure and Trustworthy ML (SatML), LLM Capture-The-Flag competition now live. Read research paper here here. We disocver that prompt-based protections to defend the leakage of 6-character secret from the system can bypassed using 4+ turn based conversations, showcasing that multi-turn Direct Prompt Injections can be very successful. The comptetion led to the collection of defense and attack dataset, published in HuggingFace </td> </tr>
[March 2024] Came out as silver medal winners at the IEEE Secure and Trustworthy ML (SatML), LLM Capture-The-Flag competition
[October 2023] Paper on "Surgical Fine-Tuening" for resource efficient task-adaptation for Language encoders accepted to EMNLP 2023. Read research paper here here </a>
[April 2023] Invited Technical Keynote Speaker at Voices of Data Science Conferenc, University of Massachusetts, Amherst </a>
[Feb 2023] Paper: Topic Segmentation for Unstructured Conversations accepted to IntelliSys Sept 2023. Read the Springer 'Intelligent Systems and Applications' book chapter here
[Jan 2023] ICLR 2023, AISTATS 2023, NIPS 2022, ICCV 2022 Reviewer
[October 2022] Paper: Topic Segmentation in the Wild: Towards Segmentation of Semi-structured & Unstructured Chats accepted to NeurIPS 2022. Read here
[Aug 2022] NLP + Reinforcement Learning lead reviewer @ Microsoft Machine Learning, AI and Data Science Conference, Fall 2022
[July 2022] Research paper on topic segmentation for noisy conversational datasets (from joint effort by University of Massachusetts at Amherst & Microsoft Research) under review at COLING 2022
[July 2022] AACL 2022 and AAAI 2023 </a> reviewer
[July 2022] Invited Speaker at WiMLDS Bay Area to present a talk about Deep Reinforcement Learning & Contextual Bandits
[May 2022] Selected as a reviewer for evaluating research paper submitted to Microsoft's Machine Learning, AI, and Data Science journal
[April 2022] Invited Speaker at ODSC East 2022 to present a talk about Quantization methods for large sequential models
[April 2022] Guest Lecturer at Carnegie Mellon University, Language Technologies Institute to present a talk about Contextual Bandits & Deep Reinforcemment Learning
[February 2022] Semi-finalist at Open AI Codex challenge (selected project: Integrating Open AI to databases to accelerate database quering process for non-developers)
[Oct 2021] Defended Ph.D. thesis on the use of deep learning and stochastic methods to detect anomalous checkpoints in sequential electricity demand profiles.
[Jun 2021] Selected to be part of the high coveted (0.1% acceptance rate) 2-year Applied ML research rotational prgram: Microsoft AI Development Acceleration Program (MAIDAP) under the Office of Applied Research as an Ph.D. Applied Scientist, post graduating from CMU.
[Jun 2021] Research paper on using deep learning methods to back-forecast electricity demand accepted at <a href => ICML</a> 2021
[Jan 2021]</strong></td> Invited to be a part of the Jane Street Ph.D. Symposium and present doctoral research related to large scale time series analysis using advanced machine learning methods
[Jun 2020] Presented research poster at WindEurope Technology Workshop 2020, Naples (virtual owing to Covid restrictions) on statistical analysis to determine changing variance of wind speeds due to climatological impacts (scenario based on RCP 6, and RCP 8.5)
[May 2020] Graduated with second Master of Science degree from Carnegie Mellon University in Engineering and Public Policy
[Apr 2020] Inducted as a part of research team at Asset Lab, University of Michigan, Ann Arbor to use stochastic methods to assess future grid reliability and also leverage advanced machine learning methods to discover patterns in large time series data with Prof. Michael Craig.
[May 2019] Selected to present first year doctoral work at the prestigious American Wind Energy Association (AWEA) Clean Power Expo and Conferece
[Mar 2019] Finalist at national level Yale Graduate Consulting Case competition.
[Aug 2018] Started serving as the Vice President of CMU Graduate Consulting Club
[Mar 2018] Elected to receive Dean's fellowship to support doctoral studies.
[Jan 2018] Awarded doctoral fellowship by CMU Portugal, to work on the +Atlantic project - a consortium of 8 cross-functional research and academic institutions from Portugal working towards assessing the technological and economical impacts of developing energy industries in the Atlantic Ocean.
[Dec 2017] Graduated with a Master of Science degree in Civil and Environmental Engineering, and started working as a Research Intern with Prof. H. Scott Matthews on ecnomic analysis and risk analysis for Deep-sea mining and Offshore wind farm infrastruture in the Atlantic region.
[Mar 2017] Received Merit Scholarship from Carnegie Mellon to support master's studies in the academic year 2017-2018.
[Sep 2016] Started serving as CMU Graduate Student Advisory Comittee Representative - a intiative to help bridge the gap between graduate students and the school administration.
[Jan 2016] Received Civil and Environmental Engineering department scholarship to support master's studies for the academic year 2016-2017. Scholarship resulted in 25% tuition waiver.