Soumedhik Bharati

Selected for a research internship at the Natural Language and Information Processing (NLIP) Lab within the Department of Computer Science and Engineering at IIT Hyderabad. Working under the mentorship of Dr. Maunendra Sankar Desarkar, my focus involves: - Advancing research in Natural Language Processing (NLP) and Information Retrieval (IR). - Developing scalable algorithms and models for real-world applications. - Collaborating with a dynamic research group dedicated to pushing the frontiers of applied AI and Indic language processing.

Xu Lab - Computational Biology Department (Advised by Dr Min Xu) Developing novel deep learning architectures for the analysis of cryo-electron tomography (cryo-ET) data. Focused on architecting generalizable segmentation and representation learning models to elucidate subcellular structures and macromolecular complexes in their native state.

Center of Computational Data Science (Advised by Prof. Pabitra Mitra) Working on improving information retrieval systems by integrating large language models in a three-stage document retrieval system inspired from WAND and TDPart. The goal is to make re-ranking algorithms more efficient and scalable without losing accuracy or relevance.

Selected as a Course Instructor to lead the Faculty Development Program (FDP) and specialized training sessions for professors and academic staff. 1. Curriculum Design & Pedagogy: Designing and delivering an advanced AI/ML curriculum tailored for academia, bridging the gap between theoretical foundations and modern industry applications. 2. Faculty Upskilling: conducting intensive sessions on Deep Learning architectures, Large Language Models (LLMs), and Generative AI to empower faculty members with cutting-edge research tools. 3. Technical Instruction: Covering topics ranging from foundational Neural Networks to advanced deployment strategies (Quantization, Distillation), enabling professors to integrate these concepts into their own research and teaching modules. 4. Workshops & Hands-on Labs: Facilitating practical coding labs and project-based learning to foster a deep, intuitive understanding of complex AI systems.

Leading core ML and AI strategy at a Polsky Center Build & Launch–backed data startup working at the intersection of voter intelligence and large-scale media analysis. Architected production-grade ML systems including an 800+ feature voter behavior modeling pipeline and advanced retrieval–reasoning architectures for news intelligence. Own technical roadmap, model design, deployment strategy, and team execution across research and production.

Designated to Exalt Data & Strategic Advisory. Responsible for the end-to-end development and deployment of AI-driven products at a strategic data advisory firm connecting officials with US voters. Work includes applying machine learning and generative AI to improve contextual relevance in high-traffic news summarization, enhancing model performance through parameter-efficient fine-tuning, and reducing model size for edge deployment using quantization and distillation. Tasks also involve designing and delivering production-ready AI solutions, managing internal teams, and collaborating with stakeholders to develop effective systems. Stay consistently engaged with the latest advancements in ML and generative AI to introduce novel techniques that support the company’s data and AI initiatives.

Providing personalized guidance and technical mentorship to students and team members in AI and machine learning projects. Supporting skill development, project milestones, and knowledge sharing within the AI/ML community. Facilitating workshops, seminars, and collaborative learning initiatives to foster growth and innovation.

The club aims to promote AIML research and development among students and faculty, and to create a community where people can learn and collaborate on AIML projects. The team members are responsible for organizing and managing the club's activities, such as workshops, seminars, and hackathons. They also work on their own AIML projects, and publish papers and blog posts about their research.

Led the end-to-end design and execution of sophisticated deep learning and hybrid modeling projects spanning EEG-based emotion decoding, genomic classification, automated essay scoring, and autonomous real-time detection systems. Directed the integration of hierarchical transformers with positional embeddings and cross-attention fusion, resulting in near-perfect affective state analysis. Oversaw the creation of memory-optimized CNN–BiLSTM hybrids reducing parameter complexity significantly while maintaining high accuracy. Managed development of attention-driven architectures for scoring tasks, reinforcement learning agents for workforce training optimization under budget constraints, and advanced UNet variants incorporating state-space layers for image dehazing. Directed predictive modeling for precise medical gait analysis and autonomous drone-based detection and neutralization systems, achieving state-of-the-art performance across diverse applications.

Played a key role in implementing and optimizing advanced neural network models, including hybrid architectures combining convolutional and recurrent layers, to tackle complex classification challenges in biomedical and genomic data. Contributed to the development of attention-enhanced sequence models and memory-efficient encoding schemes that significantly improved accuracy and computational efficiency. Assisted in designing reinforcement learning frameworks and state-space models for real-time adaptive systems, while supporting data preprocessing, model evaluation, and result documentation in a collaborative research environment.

Responsible for the design, development, and optimization of GRIT (Geometric Reprojection Instruction Tuning), a novel instruction-tuning framework that fine-tunes just 0.997% of LLM parameters while outperforming full fine-tuning and LoRA on standard benchmarks (BLEU ↑26.2, ROUGE-L ↑34.9 on Alpaca). Work includes fine-tuning multiple LLM architectures, including GPT-2 (355M), LLaMA-3B, and Mistral-7B, achieving up to 30% parameter savings without loss of performance, and reducing compute and memory usage by 40% for cost-efficient deployment. Conducted comprehensive ablation studies against LoRA, QLoRA, and AdaLoRA, validating consistent improvements of 2.1 BLEU and 2.5 ROUGE-L with GRIT’s geometry-aware updates. Maintain a strong focus on parameter-efficient learning strategies and benchmarking to advance scalable, high-performance LLM fine-tuning methods.

Subjects: Physics, Chemistry, Mathematics, Biology, Hindi, English Language, English Literature, History and Civics, Geography and Computer Science.