Pixelbin.ioDevelopment and deployment of Generative AI models has been a primary focus. Specializing in diverse areas, such as e-commerce product tagging, segmentation, image matting, text-to-image generation, and other Generative AI applications. An essential aspect of the contribution involves ensuring the seamless deployment of these models on cloud platforms. Emphasizing critical factors like scalability, reliability, and accessibility for end-users has been integral to delivering impactful solutions.ImpetusDriving innovation in fashion design tech with the Impetus project, where we’re building a next-gen AI platform tailored to the fashion industry. By integrating cutting-edge generative AI models like Flux, Stable Diffusion, and LLMs, we’re creating dynamic solutions for generation, variation, and trend forecasting. Leading the charge in developing agentic systems that automate design workflows, paving the way for seamless, intelligent design processes.

Retail AnalyticsDeveloped and optimized an end-to-end real-time object detection and analytics engine in C++ for retail and commercial stores. Trained object detection models like YOLO with PyTorch and optimized it for specific hardware with Nvidia TensorRT, achieving a throughput of 225 frames per second (FPS) on Nvidia T4 GPU and an overall accuracy of 95% on footfall analytics. Partnered with Fortune 500 companies in the retail space to deploy the solution on countless locations across the globe.Automated License Plate RecognitionImplemented real-time Automated License Plate Recognition system in C++. Trained and optimized models with PyTorch and TensorRT, resulting in a throughput of 50 frames per second (FPS) on edge devices like Nvidia Jetson TX2 NX and overall accuracy of 90% on two-wheelers and 95% on four-wheelers. Solution is used for road traffic law enforcement and specialised lane access for buses in Bus Rapid Transit (BRT) lanes in Pune city.

Computer Vision PlatformTrained and deployed models for automating power pole inspections utilizing Tensorflow, OpenCV, and other image processing tools for semantic and instance segmentation. Reduced model latency by 50% without compromising accuracy.Intelligent Automation TestingDeveloped a Q-Learning based agent for automating website testing. Eliminated the need for manual test-case creation, resulting in a more efficient website roll-out process. Seamlessly integrated with back-end frameworks such as Flask and FastAPI.

DataSwitchUsed early version of GPT-2 to create a Natural Language Interface for Databases. Developed an end-to-end application using Tensorflow and RASA, resulting in a web-service capable of formulating SQL Queries by extracting intent and entities from a user input.