Developed and deployed Generative AI applications using Custom fine-tuned Phi-3 multimodal models and Azure OpenAI LLMs, integrating computer vision techniques and meta-prompt engineering to significantly improve transcript text extraction accuracy and efficiency.
•
Implemented Retrieval-Augmented Generation (RAG) frameworks to combine LLM capabilities with external knowledge retrieval, enhancing the precision of text extraction from large-scale transcripts.
•
Designed and optimized robust data pipelines for processing and analyzing high-volume data, leveraging state-of-the-art machine learning models and cloud-based AI solutions to ensure scalability and performance.
•
Architected and built the entire backend stack to streamline and automate workflows, reducing processing latency and enhancing accuracy through seamless integration of multimodal AI, machine learning models, and advanced CV techniques.
Designed and developed scalable, high-performance applications supporting critical business processes for a retail platform with 1,500+ stores across the U.S., significantly improving system reliability and scalability.
•
Optimized transaction workflows by developing efficient APIs, resulting in a 70% reduction in transaction costs and enhancing system responsiveness.
•
Built a transfer management system using TypeScript and React to streamline and automate transfer tracking, enhancing operational accuracy and productivity.
•
Implemented dynamic, machine learning-driven pricing models, leading to a 15% revenue increase in targeted stores through optimized pricing strategies.
•
Developed and deployed computational models to forecast inventory demand, achieving a 30% reduction in excess inventory and improving overall supply chain efficiency.
Leveraged HRNet to deploy a pose detection model for soft labeling, propelling pose detection performance by 4%.
•
Developed new evaluation methodologies and metrics for human feature extraction and autonomous pose predictions in retail stores to reduce labeling costs by identifying production failures.
•
Designed and created a workflow using optical flow to identify false positives and negatives in the pose model predictions.
•
Stabilized model pipeline by reducing jitter from model outputs through filtering techniques.
Developed models using techniques like object detection and semantic segmentation for identifying and segmenting objects from street view images to aid in grid layout planning.
•
Achieved 94% accuracy with a novel architecture of deep convolution neural networks for the classification of Satellite images and Grid maps which increased the product performance by 18% in terms of classification.
•
Improved model performance by 8% using transfer learning with various CNN models for Satellite image classification.
Amplified decision support by working on data cleaning, analysis, prediction and classification of data using machine learning models like Regression, Decision trees, K-NN, Random Forest, XGBoost, K-means and Neural Networks.
•
Collaborated with senior engineers for the deployment of ML models along with multiple features for data visualization.
