Divyanshu Aggarwal

Building efficient language-model capabilities for product applications, from adaptation and training-data design through large-scale evaluation.

• Finetuning large and small language models to improve application intelligence across new tasks and domains.
• Exploring on-policy distillation of frontier models into cost-efficient, purpose-built small language models.
• Supporting a strong research culture through paper curation, reading groups, and technical presentations.

Researched how pretrained language models can gain multilingual capabilities efficiently without sacrificing existing performance.

• Developed active-forgetting and modular-learning approaches for cross-lingual transfer and language adaptation.
• Studied catastrophic forgetting, multilingual parameter-efficient finetuning, and broad multilingual evaluation.
• Ran distributed pretraining and post-training experiments on clusters spanning more than 100 GPUs.

Applied language models to customer-care and text-understanding workflows while building scalable training-data systems.

• Finetuned LLaMA models for response generation and built text-classification workflows with DeBERTa and RoBERTa.
• Combined abstractive summarization and classification for high-precision analysis of call-log data.
• Designed data curation and preprocessing pipelines on petabyte-scale Hadoop infrastructure.