cv

Bachelor

Indian Institute of Information Technology Kottayam

Computer Science (B.Tech with Honours)

Independent Researcher

Language Technology Group, University of Hamburg

Conducting research on actantial models, label projection, and LLM creativity.

• Implemented pipelines to automatically apply the actantial narrative model using LLMs.
• Designed LabelPigeon, a novel label projection technique utilized for low resource cross-lingual transfer.
• Conducted evaluations on LabelPigeon with up to a +39.9 score improvement on downstream tasks, with corresponding paper submitted to **ACL 2026**.
• Quantifying the creativity of LLMs via a novel narrative similarity evaluation scheme.

Research Intern

Language Technology Group, University of Hamburg

Conducted research on embedding models and large language model (LLM) representations.

• Conducted a general investigation on embedding models and LLM representations, developing an optimal few-shot fine-tuning regime for topic modeling.
• Devised L3Prune, a novel pruning procedure for LLM-based embedding models reducing model size by 21% with a negligible performance drop, with corresponding paper published in **RepL4NLP 2025**.

Machine Learning Intern

Institute of Human Resource Development

Developed and deployed machine learning models for speech recognition, speaker identification, and face recognition.

• Worked with the Kerala Police Intelligence Department, leading a team of 15 to develop AI systems for policing.
• Trained and deployed a state-of-the-art automatic speech recognition (ASR) model for Malayalam, along with a speaker extraction and identification system.
• Designed and implemented a Malayalam news extraction system utilizing optical character recognition.
• Developed a scalable face recognition system optimized for fast inference.

Data Science Intern

Institute of Human Resource Development

Optimized web infrastructure and data-driven decision making.

• Rebuilt web stack to streamline employee workflows, significantly improving productivity.
• Created visualizations and reports to communicate insights to stakeholders and government agencies.
• Optimized purchasing decisions based on statistical modeling.

Communications Lead

Led outreach, website development, and advocacy for marginalized communities.

• Organized events, campaigns, and workshops for the queer community in Kerala, as well as advocacy workshops and awareness programs on trans rights across universities and government institutions.
• Developed and launched a public website, significantly expanding international outreach and visibility.
• Served as the main liaison with donors, institutions, and affiliates, securing over **$3M** in grants.
• Taught math and programming to marginalized children as part of the Life Skill Education Summer Program.

Advocating for Character Error Rate in Automatic Speech Recognition

• Documented the shortcomings of the commonly used word error rate metric for multilingual evaluation in collaboration with Dr. Jesin James.
• Collected human preference data in 3 languages and calculated metric correlations, providing experimental evidence in favour of Character Error Rate.
• Wrote a paper arguing our position, accepted at **NAACL 2025**.

Fisher Mask Nodes for Model Merging

• Devised a novel and compute-efficient model merging algorithm under Dr. Suchithra M S.
• Evaluated performance of our method across various models in the BERT family, with performance improvement of **+6.5%** and a speedup between 57.4x and 321.7x.
• Documented research outcomes in an academic paper, published in **LREC-COLING 2024**.

ICFOSS Malayalam Speech Corpus

• Collaborated on IMaSC - The ICFOSS Malayalam Speech Corpus, a 50-hour text-to-speech dataset.
• Supervised data collection, speaker recording, and quality control.
• Trained and evaluated multiple models, achieving an average MOS score of 4.50.
• Wrote a research paper compiling our results, accepted at **LREC 2026**.

Whisper Malayalam

• Trained ASR models on Malayalam speech data as part of the Huggingface Whisper Fine-Tune Community Sprint.
• Elevated the medium-sized model's performance to the top of the leaderboard, making it the state-of-the-art solution for Malayalam ASR as evidenced by **30k downloads** and counting.

DAAD WISE Scholarship

German Academic Exchange Service

Just Use XML: Revisiting Joint Translation and Label Projection.

arXiv Preprint, Submitted to ACL 2026

Large Language Models Are Overparameterized Text Encoders

Proceedings of the 10th Workshop on Representation Learning for NLP (RepL4NLP-2025) (pp. 170-184)

Advocating Character Error Rate for Multilingual ASR Evaluation

Findings of the Association for Computational Linguistics: NAACL 2025 (pp. 4926-4935)

Fisher Mask Nodes for Language Model Merging

Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 7349–7355)

IMaSC -- ICFOSS Malayalam Speech Corpus

arXiv Preprint, Accepted in LREC 2026

Performance Enhancement of Deep Neural Network Based Automatic Voice Disorder Detection System with Data Augmentation — A Case Study

Biomedical Engineering: Applications, Basis and Communications, Vol. 35

Memory Based Speech Duration Model using Exemplar Theoretic Approach

Proceedings of the International Conference on Artificial Intelligence & Speech Technologies (AIST 2019) (pp. 108–112)