Project: Improved Partitioned Learned Bloom FiltersProject Advisor: Prof. Anirban Dasgupta• Designed two novel algorithms, TurboPLBF and DiffPLBF, to efficiently construct Learned Bloom Filters, resulting in a smaller memory requirement for a Bloom Filter.• The time complexity of TurboPLBF is O(nk) compared to O(nklogn) of state-of-the-art FastPLBF++. DiffPLBF has similiar performance.• Evaluation on datasets like Malicious URLs Dataset shows the superior performance of… Show more Project: Improved Partitioned Learned Bloom FiltersProject Advisor: Prof. Anirban Dasgupta• Designed two novel algorithms, TurboPLBF and DiffPLBF, to efficiently construct Learned Bloom Filters, resulting in a smaller memory requirement for a Bloom Filter.• The time complexity of TurboPLBF is O(nk) compared to O(nklogn) of state-of-the-art FastPLBF++. DiffPLBF has similiar performance.• Evaluation on datasets like Malicious URLs Dataset shows the superior performance of TurboPLBF and DiffPLBF compared to existing LBF methods.Project: Coreset for Finite-Width Neural NetworksProject Advisor: Prof. Anirban Dasgupta• Created a novel approach to Coreset construction for approximating Neural Tangent Kernels (NTKs) using Kernel Herding with super-sampling.• Extended the Coreset to Neural Networks assuming that NTKs are a good approximation for Neural Networks.• Empirically evaluated the Coreset construction for Finite-Width Neural Networks showing promising results.Project: Hadamard-based FJLT using positional value learning in CountSketchProject Advisor: Prof. Anirban Dasgupta• Designed a novel approach to create a low-distortion embedding using the Hadamard-based Fast Johnson Lindenstrauss Transform (FJLT) by learning the positions and values in the sparse projection matrix.• Adapted a learning-based approach for CountSketch to learn the sparse projection matrix in FJLT.• Evaluation on datasets showed slightly better performance than existing state-of-the-art FJLT algorithms. Show less

• Creating novel alpha strategies that involve formulating mathematical expressions, code and hyperparameters to predict the future movement of financial instruments in the USA and China markets using Technical Indicators, Fundamental Ratios, News Sentiment, Financial Models, etc.

• Applied Data-centric AI methodologies to improve the robustness of a production NER model with an increase of 10%+ in accuracy.• Explored generative models along with prompt design and tuning for Large Language Models to study their application in the biomedical domain.• Co-authored a research paper on Multi-Task Learning for Biomedical NER, which has been published. It was part of a team-wide effort to decrease the latency of transformer NER models; in effect, a 20x speedup was… Show more • Applied Data-centric AI methodologies to improve the robustness of a production NER model with an increase of 10%+ in accuracy.• Explored generative models along with prompt design and tuning for Large Language Models to study their application in the biomedical domain.• Co-authored a research paper on Multi-Task Learning for Biomedical NER, which has been published. It was part of a team-wide effort to decrease the latency of transformer NER models; in effect, a 20x speedup was achieved. Show less