Projects
Dynamic Activity-Dependent Pruning (DADP)
Timeline: Mar. 2026 – Sep. 2026 | Role: First Author | Status: Under Review, ICLR 2027 (arXiv:2610.11853)
- Built a single-pass pruning method inspired by synaptic plasticity to compress neural networks efficiently.
- Discovered that pruned vision models naturally retain early sensory layers, while transformers keep late context layers.
- Pruned VGG-16 to 98.41% sparsity (63× compression) on MNIST maintaining 98.72% accuracy, and to 72.79% sparsity on CIFAR-10 with +1.70% accuracy improvement over dense baseline.
Olfactory-Inspired Sparse Combinatorial Coding
Timeline: Dec. 2025 – Jun. 2026 | Role: Independent, First Author | Preprint: arXiv:2606.21895
- Designed a biologically inspired sparse compression layer based on mammalian olfactory circuits.
- Implemented an Olfaction Module + BiLSTM-CRF architecture.
- Reduced model parameters by 44% while improving low-resource F1 score by up to 6.23% over dense baselines.
Retrieval-Augmented Large Code Generation and Evaluation (MS Thesis)
Timeline: May 2024 – May 2025 | Affiliation: Coriolis Technologies / IISER Pune
- Developed RAG system to synthesize complex 600–800-line Python code, improving zero-shot baseline accuracy from 20% to 65–70%.
- Designed an iterative generation and self-correction loop with feedback, lifting overall accuracy to 90%.
- Formulated an LLM-as-a-Judge evaluation framework to assess synthesized code generated in every iteration.
- Benchmarked hallucination dynamics and instruction-following across 12+ open-source models (1B–24B parameters) under 40 GB VRAM.
Scaling up data curation workflows using NLP
Timeline: Aug. 2023 – May 2024 | Affiliation: IISER Pune | Supervisor: Prof. Joy Merwin Monteiro
- Leveraged modern NLP techniques to develop workflows that accurately curate address data.
- A basic geocoding framework is created using Weak Supervision and Clustering methods.
Normalization analysis with VGG16 on CIFAR
Timeline: May 2023 – Oct. 2023 | Affiliation: IISER Bhopal | Supervisor: Prof. Akshay Agarwal, PhD
- Implemented VGG16-based Convolutional Neural Network (CNN) for image processing tasks.
- Evaluated various normalization techniques and identified Group Normalization as the most effective.
- Validated findings on CIFAR10 and CIFAR100, where Group Normalization achieved the fastest convergence and highest accuracy.
Era prediction based on text analysis
Timeline: Aug. 2022 – Dec. 2022 | Affiliation: IISER Pune | Supervisor: Prof. Leelavati Narlikar
- Developed a machine learning model to predict the era of a text using Markov models and log-likelihood analysis.
- Conducted text analysis on books from the 18th and 20th centuries.
- Implemented 0th and 1st-order Markov models, achieving an AUC of 83% with the first-order model.