Microsoft Research India
Research Fellow
Researched how pretrained language models can gain multilingual capabilities efficiently without sacrificing existing performance.
Sep 2023 — Jul 2025
Delhi, India
About me
I am an Applied Scientist 2 at Microsoft, where I work on adapting and distilling language models for product applications. My current work focuses on efficient post-training, synthetic training data, and large-scale evaluation for expanding model capabilities across new tasks and domains.
Previously, I was a Research Fellow at Microsoft Research India, where I was advised by Dr. Sunayana Sitaram, Dr. Satya Lokam, and Dr. Navin Goyal. I worked on multilingual parameter-efficient finetuning, continual learning and catastrophic forgetting, language adaptation and cross-lingual transfer, and multilingual benchmarking. The broader goal was to extend language-model capabilities to new languages and domains through efficient, modular methods.
I have also collaborated with Dr. Vivek Gupta, Dr. Anoop Kunchukuttan at AI4Bharat, and Dr. Ashwini Vaidya at IIT Delhion benchmark datasets, multilingual inference, semantic parsing, and adapting language models for Indian languages.
Earlier in my career, I held applied NLP roles at American Express AI Labs and Builder.ai, working on language models for risk, customer-care, recommendation, and conversational use cases. I graduated from Delhi Technological Universityin 2021.
Please feel free to reach out by email if you would like to discuss my research or potential collaborations.
Past experience
Selected applied research and data science roles before my current position.
American Express AI Labs
AI Researcher
Applied language models to customer-care and text-understanding workflows while building scalable training-data systems.
Aug 2022 — Aug 2023
Builder.ai
Data Scientist
Worked on conversational recommendation, orchestration, and intent understanding for product-building workflows.
Mar 2021 — May 2022
Research interests
My work combines algorithms, evaluation, and systems for adapting language models beyond their original languages and training distributions.
Efficient adaptation
Adapting large and small language models to new applications and domains through finetuning, parameter-efficient methods, and model distillation.
- Finetuning
- PEFT
- Distillation
Multilingual learning
Expanding language capabilities through cross-lingual transfer, modular learning, and multilingual data while preserving existing model behavior.
- Cross-lingual transfer
- Continual learning
Evaluation at scale
Understanding model quality across languages, tasks, and modalities, with reproducible benchmarks and large-scale training and evaluation systems.
- Benchmarking
- LLM evaluation
- ML systems
Selected publications
Research on multilingual adaptation, continual learning, efficient finetuning, and language-model evaluation.
Exploring Two-Phase Continual Instruction Fine-tuning for Multilingual Adaptation in Large Language Models
A two-phase instruction-finetuning approach for extending multilingual ability while retaining previously learned capabilities.
Improving Cross Lingual Transfer by Pretraining with Active Forgetting
Pretraining with active forgetting to improve transfer when decoder language models are adapted to new languages.
Improving Consistency in LLM Inference using Probabilistic Tokenization
A probabilistic tokenization strategy for improving the consistency of language-model outputs across repeated inference.
MAPLE: Multilingual Evaluation of Parameter Efficient Finetuning of Large Language Models
A systematic evaluation of parameter-efficient finetuning choices across multilingual downstream tasks.
MEGAVERSE: Benchmarking Large Language Models Across Languages, Modalities, Models and Tasks
A broad evaluation benchmark spanning languages, modalities, model families, and task types.
* Equal contribution
Selected contributions
Contributions to widely used training, adaptation, and interpretable-ML tooling.
News
This list is scrollable. Use the arrow keys or Page Up and Page Down to view older updates.
Our work on two-phase continual instruction finetuning appears in ACL 2026 Findings.
Our active-forgetting work on cross-lingual transfer appears at EMNLP 2025.
Joined Microsoft India as an Applied Scientist 2.
Our probabilistic-tokenization work appears in NAACL 2025 Findings.
Our work on probabilistic tokenization was accepted to NAACL 2025 Findings.
Released our preprint on active forgetting for cross-lingual transfer.
Released the MAPLE preprint on multilingual parameter-efficient finetuning.
Joined Microsoft Research India as a Research Fellow with Dr. Sunayana Sitaram.
Our work on inter-bilingual semantic parsing was accepted to NLP4ConvAI at ACL 2023.