Led the end-to-end development of a voice-based AI hardware device that enables children to craft interactive stories, generate digital artwork, and compose original music through natural conversation.• Architected and implemented an end-to-end voice-based interaction system leveraging Google Cloud Platform for LLM orchestration• Engineered a low-latency cloud infrastructure for streaming Text-to-Speech (TTS) audio outputs using websockets, enhancing user experience• Implemented comprehensive testing frameworks to monitor LLM accuracy and maintain prompt engineering stability• Leveraged and fine-tuned LLaMA 70B model for robust tool-calling capabilities• Optimized TTS performance using TensorRT, significantly improving end-user experience

Systems analysis and design teaching assistant.

Model Development:- Successfully developed and deployed intricate ML models for IoT-based devices, leveraging edge and cloud computing.- Developed classification models trained on 6 million+ images, predicting over 1700 classes with robust scalability on GCP (50 million+ monthly inferences). Research on Semi-Supervised Learning Models:- Spearheaded R&D efforts for an innovative semi-supervised learning model, harnessing model embeddings and vector search techniques.- Achieved an outstanding >80% reduction in time and labeling costs, transforming efficiency and minimizing reliance on manual data labeling. Optimization for Edge Devices:- Optimized models utilizing int8 quantization for edge devices (TFLite & ONNX), resulting in a notable 40% latency reduction and significant model size reduction for improved real-time processing. Image Processing and Computer Vision:- Integrated OpenCV-based image processing and advanced CV algorithms to meticulously sort images based on attributes like resolution, blur, and noise.- Significantly expedited dataset creation, contributing to substantial reductions in overall creation time. Model Deployment (MLOPS):- Orchestrated strategic deployment of cloud-based ML models across diverse global regions using Google Cloud’s Vertex AI platform.- Implemented sophisticated threading techniques, reducing deployment time by a remarkable 60%, thus enhancing operational efficiency.

Utilized LSTM and transformer architectures within the domain of Natural Language Processing (NLP) for a genre classification task based on movie summaries.- Refined model performance through hyper-parameter optimization techniques.- Accelerated real-time inference on GPUs by implementing TensorRT for enhanced NLP-based genre classification.

Worked on developing a solution to extract documents and visiting cards in animage and extract valuable information to help clients gain insights and customerdetails. The solution used a variety of computer vision, NER, and NLP, techniques,and a REST API back-end was created to integrate the solution with the application.