Jan 2022 – Dec 2023 · 1 yr 11 mos · Bengaluru, Karnataka, India

• Designed and analyzed Interactive Machine learning models for fast personalized recommendations via
• Online Collaborative Filtering algorithms. Our approach uses low rank matrix completion techniques with
• provable theoretical guarantees. Our designed algorithm is scheduled to be implemented for predicting available call-slots to provide advice for 2.2 million underprivileged expectant mothers. In simulations, our algorithm is providing a 20% reduction in cost over non-collaborative algorithms in the above application.
• Designed Robust Meta Learning Algorithms with privacy guarantees for multi-task learning/meta-learning in a novel Low Rank and Sparse framework that carefully model parameter sharing across users. Used
• fast Alternating Minimization techniques with provable guarantees. Implemented designed algorithm for
• fine-tuning pre-trained Deep Neural Network thus personalizing Google’s Email Ranking program - led to
• > 2% increase over baseline MRR obtained without any personalization (across > 100k enterprises).

Feb 2020 – May 2020 · 3 mos · Berkeley, California, United States

• Worked on PECOS and Sequence to sequence deep learning models in order to predict the queries of a user given the previous query in a prefix-free setting. I completed a detailed evaluation of both these models on random traffic data and came up with conclusive evidence to suggest that Seq2seq performed better than PECOS.
• I designed a new two-stage model that combined PECOS and Seq2seq by clustering the ELMO embedding of each word in the vocabulary and training the seq2seq model on the clusters as tokens. In the second stage, I trained a linear ranker to rank the items in each cluster.
• The new model had the best Weighted BLEU score beating the state-of-the-art pre-trained models and also obtained a comparable MRR score. Further, this model performed better than all other naive baselines.

May 2019 – Aug 2019 · 3 mos · Palo Alto

• Worked on document segmentation in order to recover Unit of Analysis (UoA) from a single page invoice. I developed an end-to-end trainable deep learning model using LSTM that learns the similarity between two bounding boxes of words present in a document.
• From the resulting fully connected weighted graph, I bypassed the necessity for the graph clustering problem (and therefore any hyperparameters) by developing a novel dynamic programming algorithm for inference. The DP algorithm was completely parameter-free and therefore domain independent.
• I also designed a fast version of this algorithm that is able to achieve highly accurate results on documents and required an approximate time of only 20 seconds.

Aug 2016 – Nov 2021 · 5 yrs 3 mos · Amherst (MA), United States

• I completed my Ph.D under the supervision of Dr. Arya Mazumdar at the University of Massachusetts Amherst. My thesis titled "Mixture Models in Machine Learning" can be found at this link https://web.cs.umass.edu/publication/docs/2022/UM-CS-PhD-2022-001.pdf

May 2015 – Jul 2015 · 2 mos · Greater Bengaluru Area

• Worked on open source codes of XTK viewer and Rest-3D server and integrated them and
• modified them to include a large number of features
• Extensive work on javascript and Node.js to build both server side and client side systems that
• interact with each other by sending and responding to requests.
• Guide:-Ashes Dhanna Ganguly

May 2014 – Jul 2014 · 2 mos

• Worked on the establishment of a Line of Sight radio link for horizontal communication using the
• modern DVB-S and DVB-S2 technology used for satellite communication.
• This work, if implemented in reality, will lead to a quasi-error free horizontal communication
• Guide: Prof. Dr. Rabindranath Bera, Member of the Indian Academy of Sciences, India
• See project Development of Communication Model