Architected and implemented an end-to-end automatic speech-to-text system (from data pipelines to model re-training) that transcribes and de-hallucinates doctor-patient conversations for the extraction of meaningful audio and linguistic signals.

Prompt-engineered and deployed LLMs such as GPT-4, GPT-3.5, Mistral-7B, Llama-3 for speaker diarization with zero-shot, few-shot and chain-of-thought learning approaches.

Fine-tuned Mistral-7B-Instruct and Llama-3 models using PEFT-based approaches, including LoRA, QLoRA, DoRA, and QDoRA, for speaker diarization pipeline, achieving state-of-the-art results.

Conducted evaluation of internal SMPA pipeline against open-source and proprietary ASR models from Google Cloud, Amazon, and Microsoft Azure and achieved SOTA results for multi-lingual recordings.

Led the design and implementation of an automated AWS Lambda-based ETL pipeline, extracting and integrating data incrementally from RAVE Medidata API into AWS Aurora.

Created 20+ data jobs with SCD2 and history tables using SQL and Matillion for Redshift.

Collaborating with data engineers and Urovant consultants to normalize pharmaceutical data.