Rohit Kumar

Research Engineer-1 - Avataar.ai

July 2024 - Present

• Built an end-to-end pipeline that automatically creates lifestyle images using Flux Model and ControlNets, making products look better and more realistic for customers.
• Modified the diffusion model’s sampling procedure to improve object reconstruction, followed by intrinsic decomposition for realistic product relighting.
• Built classification systems for low-data scenarios using pre-trained models such as CLIP, BLIP2 and Qwen2.5, enhancing accuracy in product categorization.
• Enhanced segmentation accuracy by implementing BiRefNet models while also integrating SAM with YOLO-world for complex scenes.
• Improved object detection by benchmarking various frameworks including YOLO-world, Florence, and other mechanisms.

Teaching Assistant - IISc Bangalore

Subject: Signal Processing in Practice

Jan 2024 - Apr 2024

• Integrated continual learning frameworks (L2P, DualPrompt) to mitigate catastrophic forgetting in neural networks.
• Built self-supervised models using MoCo and SimCLR for robust visual representation learning.
• Developed adaptive prompt-based learning with dynamic token expansion and attention mechanisms.

Teaching Assistant - IISc Bangalore

Subject: Digital Image Processing

Aug 2023 - Dec 2023

• Developed DFT-based frequency domain filtering for advanced image denoising and enhancement.
• Implemented SIFT and Normalized Cut for precise feature detection and image segmentation.
• Optimized deep learning models using EfficientNet-B0 with custom classifiers for efficient convergence.