Chinmay Choudhary

WORK EXPERIENCE

• Enhanced Text-to-Speech models for Indian and regional Indian accents, improving user experience. Developed a framework for objective evaluation of generated audio and increased consistency of fine-tuned models.
• Engineered fine-tuning pipelines for OpenAI Whisper model implementing modular layer freezing and knowledge distillation, and PEFT using PyTorch Lightning and HuggingFace Transformers
• Achieved Word Error Rate (WER) performance within 4% of state-of-the-art models for telephonic Indian audio transcription through custom fine-tuning approaches
• Implemented real-time streaming of OpenAI Whisper over WebSocket using NVIDIA TensorRT, achieving 200ms average response time with 10 concurrent client connections on a single GPU setup
• Developed end-to-end data collection and processing pipelines incorporating Label Studio for systematic data labeling of speech-to-text and text-to-speech training datasets

• Fine-tuned wav2vec2 models for regional multilingual automatic speech recognition, achieving a Word Error Rate (WER) score of 30%
• Implemented context-aware RAG-based question-answering chatbots with streaming capabilities
• Engineered a high-performance, streaming-enabled pipeline using various generative AI models
• Conducted comprehensive evaluations of open-source technologies from various providers

• Fine-tuned a large language model with 700 million parameters using QLoRA and PEFT
• Led efforts to improve model's human alignment and reduce toxicity using RLHF
• Implemented GCP Workbench AI for model hosting and training
• Created ensemble neural network regressors achieving MSE value of 0.13
• Utilized Vision Transformers for document classification with 97% accuracy

• Developed scalable RESTful API solutions using Flask and Python
• Managed deployment of ML projects using Docker, Jenkins, and AWS
• Led team in creating NLP classification model with 90% f1 score

PROJECT EXPERIENCE

• Implemented dynamic system to update cricket team ratings post-match
• Analyzed IPL teams' Elo ratings and performance metrics
• Created visual representations of Elo rating progressions

• Performed TimeSeries analysis with ADF Fuller test
• Developed RNN models with LSTM for feature extraction
• Achieved 0.8% Mean Percentage Error in multivariate model

EDUCATION