AI Research Engineer at Logituit
I am an AI Research Engineer and NLP researcher working at the intersection of language models, efficient learning, and human-centred AI. As a member of the Lingo Research Group at IIT Gandhinagar, I conducted research examining how limited annotation effort, training tokens, and examples can be allocated more effectively. My broader interests include instruction tuning, multilingual learning, personalized recommendation systems, retrieval-augmented generation, evaluation, data quality, and human feedback.
AI Research Engineer · Recommendation Systems
A lightweight record of what I am working on, exploring, and planning next.
Happy, glad and grateful that Soket AI is sponsoring my travel to attend EMNLP 2026 conference at Budapest, Hungary.
TSS paper got accepted to EMNLP 2026, main conference, looking forward to attend, at Budapest, Hungary.
I have joined logituit, new place, new people, I hope everything goes well.
My research focuses on developing data-efficient, multilingual, and reliable language models. I am particularly interested in how limited data, computation, and human feedback can be used more effectively when training, adapting, and evaluating AI systems.
Budget-aware training and data selection Efficient allocation of annotation and training budgets across tasks, examples, and annotators, including the selection of high-value data for language-model adaptation.
Instruction tuning and multilingual transfer Efficient adaptation of language models through instruction tuning and parameter-efficient methods, with an emphasis on transfer between high- and low-resource languages.
AI alignment and human feedback Developing reliable and inclusive methods for aligning language models with diverse human preferences while using human feedback efficiently.
Evaluation, annotation, and data quality Studying how annotation strategies, evaluator diversity, instance difficulty, and dataset quality influence model behaviour and evaluation reliability.
Retrieval-augmented generation and recommendation systems Combining language models with retrieval, structured knowledge, user behaviour, and content representations to build grounded conversational and recommendation systems.
More broadly, I am interested in building AI systems that are efficient, multilingual, trustworthy, and responsive to the people and contexts in which they are used.
My doctoral research at IIT Gandhinagar studied budget-aware decision making for NLP: how to use limited annotation effort, training tokens, and examples more effectively. The broader motivation is that better performance does not always come from simply collecting more data or training for longer; it also depends on deciding where the available resources should be spent.
Annotation: I studied the trade-off between collecting more labels for the same example and labeling a larger number of examples. Using a controlled multi-annotator simulation framework, I examined how annotation depth, disagreement, and instance difficulty affect model performance under a fixed annotation budget. The results show that collecting multiple labels is not always beneficial; its value depends on which examples receive additional annotations and how difficult or ambiguous those examples are. This work appeared in EMNLP 2023.
Training: I studied how a fixed token budget should be distributed across different task types during instruction tuning. Instead of relying on uniform or dataset-size-based sampling, I developed ADAPT, a validation-driven method that learns task proportions during training. ADAPT treats task allocation as a dynamic decision: it uses validation feedback to shift more of the available training budget toward tasks that are useful for downstream generalization.
Selection: I studied how to identify instruction-tuning examples that provide a meaningful instruction-following signal before fine-tuning. I developed Task-Specificity Score (TSS), which measures how strongly an output depends on the given instruction, conditioned on the input. Its extension, TSS++, combines instruction specificity with harder counterfactual instructions and data-quality signals. This allows models to train on smaller but more useful subsets of instruction data.
Together, these projects show that annotation, training, and data selection should be treated as allocation problems rather than simple scaling problems. My thesis develops methods for making these decisions more systematically under realistic resource constraints.
Alongside my thesis, I also worked on understanding how machine-learning models are shared, documented, and adopted in public model repositories. In Model Hubs and Beyond, we studied model popularity, performance, and documentation practices on Hugging Face. The work examined around 500 sentiment-analysis models and more than 80,000 human annotations, showing that model popularity is not determined by performance alone and that documentation quality, visibility, and platform-level factors shape which models receive attention and reuse. This project broadened my research from efficient model training to the wider infrastructure and social processes through which models are communicated and adopted.
Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pp. 44255–44267.
Proceedings of the 20th International Workshop on Semantic Evaluation, pp. 2991–3001. DOI
arXiv preprint arXiv:2605.00817. DOI
arXiv preprint arXiv:2512.04555. DOI
Proceedings of the Nineteenth International AAAI Conference on Web and Social Media, pp. 965–993. DOI
Findings of the Association for Computational Linguistics: EMNLP 2023, pp. 1371–1388. DOI
Proceedings of the ACM/IEEE Joint Conference on Digital Libraries in 2020, pp. 539–540. DOI
My teaching experience spans data-centric computing, deep learning, natural language processing, and database systems. I have supported large student cohorts through lectures, tutorials, assignments, evaluations, project mentoring, and doubt-solving sessions.
Served as a Graduate Teaching Fellow for Data-Centric Computing, a foundational course introducing students to computational thinking through data. I supported topics such as data representation, programming fundamentals, data processing, visualization, basic statistical reasoning, and problem-solving with real datasets. My role included conducting tutorials, assisting with assignments and assessments, clarifying programming concepts, and helping students connect computational methods with practical data-driven applications.
Served as a Graduate Teaching Fellow for a Deep Learning course with approximately 300 students. Supported the teaching of neural networks, optimization, convolutional networks, sequence models, attention, transformers, and modern language models. I contributed to tutorials, programming assignments, evaluation, project guidance, and student mentoring, with an emphasis on connecting mathematical foundations to practical implementation using modern deep learning frameworks.
Assisted with courses covering language modeling, text classification, sequence labeling, embeddings, attention, transformers, multilingual NLP, and evaluation. Guided students through programming assignments and projects involving real-world datasets and transformer-based models.
Supported instruction in relational modeling, SQL, normalization, indexing, query processing, transactions, and database design. Conducted tutorials and helped students connect database concepts with practical software and data-system development.
I aim to make technical concepts accessible without oversimplifying them. My teaching combines conceptual clarity, mathematical intuition, implementation, and real-world examples, while encouraging students to analyze failures, question assumptions, and build reliable systems.