Jan 2025 - Ongoing

Part of the Input Experience NLP team building language models across input experiences and writing tools for Apple Intelligence.

May 2024 - Aug 2024

• Investigated data curation and augmentation algorithms to improve LLM pretraining efficiency through high‑quality inputs.

• Explored data diversification through difficulty and density‑based selection in the representational embedding space. Pretraining on this curated data improves training efficiency by 28.6% on large‑scale text corpora.

• Synthetically rephrased raw corpora into high‑quality learnable formats reducing average perplexity on Pile by 31% and providing 2.24x training speedup compared to equivalent raw‑only data

July 2022 - July 2023

• Developed a multimodal late-fusion method using text and past activity signals to detect content integrity violations in Virtual Reality. 

• Exploring synthetic generation and few-shot metric learning paradigms that can be extended over text and image to further improve performance for detecting content integrity violations in low-resource settings.

Jan 2022 - May 2022

• Developed a method to solve Raven’s progressive matrices inspired from the Neural Algorithmic Reasoning aprroach.

• Worked on auto-encoders trained with the objective of mapping symbolic and image representations to the same latent space.

• Trained neural predicates and used them to search the combination of rules that satisfy the problem instance

Jan 2022 - March 2022

• Part of the Moderation Automation and Relevance Systems (MARS) team, which ensures safe advertising. Worked on improving model performance for automated moderation to detect perceptual quality and letter box defects in video ads.

• Built and fine-tuned 3D ResNet based video models for video quality assessment experiments. Performed inference using SoTA models for image quality assessment and scaled it for video quality task. 

• Trained VGG-13 to detect frames with letter-pillar box defects, this improved precision to 93.80% and recall to 62.20%

Jan 2022 - April 2022

• Worked under the supervision of Dr. Camilo Rojas. Part of the Machine Learning team of Project Us which aims to help people develop their sense of empathy. It uses AI to analyze interlocutors’ signals (e.g., speech, transcript, facial expressions) and performs emotional content analyses, which is interfaced back in real-time.

• Worked on improving model inference framework to achieve higher performance, resource efficiency, scalability and draw more insights to be provided on the dashboard

Jan 2022 - April 2022

• Worked under the supervision of Dr. Gautam Shroff and Dr. Ashwin Srinivasan.  Formulated and applied the idea of Neural Algorithmic reasoning on RAVENs progressive matrices. Observed near human performance for 4/7 configurations.

• Paper for the same accepted at NeSy-IJCLR 2022

Jan 2021 - Apr 2021  and Aug 2021 - Dec 2021

• Worked on techniques to improve performance of deep learning models across various modalities using interpolative data augmentation method - Mixup. Built saliency based and hyperbolic space variants of mixup that can  improve performance for various downstream tasks. 

• Stock volatility forecasting using multimodal - transcript and speech data from earnings calls. Identifying and analysing demographic bias in high performing architectures for this task.

 May 2021 - July 2021

• Worked on the project which aims to enhance the Gmail Search functionality. 

• Built a flume pipeline to replay queries and analyse the aggregate differences in the proposed query parsing stack. Differences can then be fixed to achieve parity and enhance search quality.

 May 2020 - July 2020

• Worked on the project Carbon Dashboard which provides granular insights and analysis of carbon emissions across buildings. 

• Developed a Machine Learning Time Series SARIMAX Model to predict people attendance in offices.  This allows for better planning, reducing resource wastage and carbon emission.

 May 2020 - June 2020

• Built deep learning models for Structural Health Monitoring-Damage Classification using inter-storey drift ratio.

• Trained CNNs on simulated accelerometer signals for classifying structural state based on damage. This can help in addressing vulnerable structures in the aftermath of a natural event. [Repo]