May 2022 – Jul 2022 · 2 mos · Mumbai, Maharashtra, India

• Worked on building an end-to-end NLP-driven meeting summarisation system for JioMeet, leveraging transformer architectures to automate post-meeting insights and reduce manual note-taking for users.
• Key Contributions:
• >Designed an extractive summarisation pipeline using TF-IDF ranking, sentence embeddings, and graph-based scoring, achieving ~33% compression while retaining core semantic content.
• >Implemented BERT-based models (BERT & DistilBERT) to improve contextual sentence ranking and summary relevance, significantly outperforming classical NLP baselines.
• >Experimented with abstractive summarisation using transformer seq-to-seq architectures, evaluating models such as BART/T5 to produce more human-like summaries with improved coherence.
• >Developed a complete speech-to-text → transcription → NLP summarization workflow:
• Extracted audio from JioMeet recordings
• Applied Python-based STT APIs for transcript generation
• Preprocessed text (sentence segmentation, tokenisation, embeddings)
• Generated extractive + abstractive summaries
• > Optimised NLP inference by batching tokenisation, reducing transformer forward-pass overhead, and improving preprocessing throughput.
• >Conducted evaluation using ROUGE metrics to benchmark model performance and guide iterative improvements.

Mar 2022 – Jun 2022 · 3 mos · Gurugram, Haryana, India