experience

Education

GPA 3.87/4.0.

Research Experience

• Translated the English BabyLM corpus into Hindi and Telugu with IndicTrans2 and trained GPT-BERT and GPT-Wee models (about 98M parameters) on 4× A100 GPUs with Hugging Face Accelerate.
• Published in the Proceedings of the Second BabyLM Workshop (ACL, 2026); corpora released on Hugging Face.

• Trained five GPT-2 models from scratch: four monolingual models with 32K tokenizers and one joint model with a shared 64K tokenizer, on cleaned CC-100, Wikipedia, and Samanantar data.
• Monolingual models outperformed mGPT on sentiment classification and NER, and their tokenizers were more efficient in all four languages.

• Built a longitudinal evaluation of 8 release transitions across 4 providers and 7 bias benchmarks (12,789 filtered responses), measuring sentiment, toxicity, and stereotype drift.

• Studying how regional and dialect variation affects tokenization, representations, and transfer in language models.

Professional Experience

• LLM systems for network intelligence: anomaly analysis, agentic AI, and MCP.

• RAG pipelines, LLM applications, and ML infrastructure on Azure.

• Backend APIs and data reliability on AWS Lambda and PostgreSQL.

• Backend services and microservices in Python and Java on AWS.