Jun 2024 – Aug 2024 · 2 mos

• Developed novel machine learning algorithms to detect instance-wise spurious correlations between features and class labels in multimodal datasets, using robust metrics to rank instances and automate the curation of high-quality datasets for model training.

Jan 2024 – Dec 2024 · 11 mos · New York, United States · On-site

• Researched on controlled text generation of LLMs such as Llama, GPT-3.5, and GPT-4 at the NLP Research Lab for humor generation.
• Focused on guiding LLMs to generate and suppress content according to user specifications in embedding and probability spaces.
• Improved humor generation of LLMs using various RL techniques like PPO & DPO.
• Co-authored a research paper on leveraging LLMs to create synthetic datasets for humor research.

Sep 2023 – Dec 2023 · 3 mos · New York, United States · On-site

• Worked on Low-Resource Machine Translation to translate Prakrit, an ancient Indian language, into colloquial languages like English, Hindi, and Gujarati, significantly improving the BLEU score for English translation.
• Built a conversational agent using Llama, RAGs, Langchain, and FAISS to handle multilingual queries and responses.
• Co-authored a paper analyzing the impact of fine-tuning large multilingual models on translation quality for semantically similar languages, highlighting both effectiveness and limitations.

Aug 2022 – May 2023 · 9 mos · Kharagpur, West Bengal, India · On-site

• Conducted research on NLP techniques to enhance Machine Translation for low-resource Indian languages.
• Improved translation accuracy of NER tokens and inflectional verb forms using geometrically-aware bilingual word embeddings and POS tagger Trankit in a Template-based format optimal for training Transformer models.
• Explored optimal synergies by altering loss functions to incorporate continual learning on numerous neural networks such as Transformers, BERT, and LeCA using Fairseq, PyTorch, and Python.

May 2022 – Jul 2022 · 2 mos · Hyderabad, Telangana, India · On-site

• Devised an efficient data monitoring system to assess Kafka brokers' health and data pipeline's bandwidth leveraging out-of-band Prometheus metrics in Scala to gauge latency accurately and ensure better customer experience through Grafana dashboards.
• Deployed monitoring system into production using Jenkins pipeline and contributed to the Kafka GitHub repository, providing enhanced visibility into data throughput at user level, and alert customers of potential pipeline downtime.