Apr 2023 – Jan 2024 · 9 mos · Remote

• Supernormal captures meetings and uses transcriptions to automatically generates notes, summaries, and action items using large language models (LLMs)
• I engineer prompts and fine-tune LLMs for note-quality enhancement and develop metrics for quality evaluation

Jun 2022 – Aug 2022 · 2 mos · Menlo Park, California, United States

• Worked with the Creative Delivery Team on ad enhancement through augmented reality effects
• Built pipelines to identify impactful AR effects, and made transformation infra guardrails configurable

Jan 2021 – Jul 2023 · 2 yrs 6 mos · Pittsburgh, Pennsylvania, United States

• 1. Multilingual Classification of Environment-related News using LLMs
• Collaborated with the World Wide Fund for Nature (WWF) to identify multilingual news content related to environmental conservation
• Designed a multi-step prompt pipeline to perform multilingual classification for low-resource languages, using in-context learning, chain-of-thought reasoning, and self-reflection on LLMs
• Achieved state-of-the-art performance using <6% of the training data required by existing models
• 2. Evaluation of Generated Text using In-context Learning (ACL Findings '23)
• Designed a framework to use large language models as evaluators of artificially generated text using in-context learning
• Engineered a prompt that achieves state-of-the-art evaluation of text summarization without requiring the evaluation model to be fine-tuned on large datasets.
• Link: https://aclanthology.org/2023.findings-acl.537.pdf
• 3. A Mixed-method Reflexive Analysis of the Fairness, Accountability, and Transparency Conference (FAccT '22)
• Identified areas of focus in the fair machine learning domain through community detection in
• fair-ML citation networks and a metastudy of the FAccT conference.
• Uncovered the moral underpinnings of FAccT against the framework provided by the Moral
• Foundations Theory.
• Link: https://dl.acm.org/doi/pdf/10.1145/3531146.3533107

Jul 2019 – Dec 2020 · 1 yr 5 mos · Bangalore Urban, Karnataka, India

• Worked with the Natural Language Understanding division of the Voice Intelligence team on the Development of Bixby–Samsung Electronics' personal assistant.
• Implemented an NLU engine to run under on-device resource constraints, thereby reducing server costs and improving offline performance.

Jan 2019 – May 2019 · 4 mos · Zurich, Switzerland

• Worked at the Language and Space Lab on undergraduate thesis titled "Copy mechanisms for Upstream Text Processing".
• Designed an encoder-decoder model with soft attention and a character copying mechanism to
• improve low-resource performance.
• Applied the architecture for normalization of the Swiss-German language and for morphological
• segmentation of the English, German, and Indonesian languages.

May 2018 – Jul 2018 · 2 mos · Aix-En-Provence Area, France

• Worked within the framework of the ACORFORMED project, which aims to develop an embodied conversational agent (ECA) to train doctors to break bad news.
• Developed a model to estimate the levels of presence and co-presence experienced by a participant during his/her interaction with an embodied conversational agent in a virtual reality environment on the basis of verbal and non-verbal features.
• Built random forest and support vector machine models to estimate presence and co-presence from the features obtained.

May 2017 – Jul 2017 · 2 mos · Greater Delhi Area

• Built a prototype image classifier aimed at identifying frequently encountered objects on roads as a component of an autonomous car development project.
• Developed a CNN based model by performing transfer learning using pre-trained weights from the VGG-16 Network.