CV

Master Thesis Researcher – Improved Driver Distraction Detection using Self-Supervised Learning

Fraunhofer IOSB

Investigated cross-modality generalization using self-supervised vision transformers for driver distraction detection.

• Innovated the Clustered Feature Weighting (CFW) algorithm using HDBSCAN for long-tailed distribution.
• Achieved 7.17% performance improvement over supervised baselines using DINOv2-pretrained ViT-B/14.
• Used tools like Ray, PyTorch, Scikit-learn, OpenCV, and Git for reproducible experimentation.

Research Assistant (KARLI and SALSA Projects)

Fraunhofer IOSB

Conducted AI research for human-machine interaction and autonomous driving projects (KARLI, SALSA).

• Developed an LLM-based chatbot using RAG and RAPTOR for scalable retrieval in knowledge-intensive settings.
• Designed the Clustered Feature Weighting (CFW) algorithm to mitigate class imbalance in multimodal datasets.
• Contributed to VoxelNeXt deployment for LiDAR-based 3D object detection on the Pandaset dataset.
• Led technical demos of a Level 3 Mercedes-Benz Occupant Monitoring System for stakeholders.
• Worked with Python, PyTorch, Ray, FAISS, RAPTOR, LangChain, Docker, Streamlit, and AWS.

Conference Reviewer

IEEE ITSC 2025 (International Conference on Intelligent Transportation Systems)

Served as a peer reviewer for the IEEE ITSC 2025 conference in the domains of machine learning and computer vision.

• Reviewed submissions in unsupervised learning, segmentation, and safe motion generation.
• Focused on evaluating methods in ML, CV, and intelligent systems for transportation and autonomous driving.

M.Sc. Mechatronics

University of Siegen

Department of Electrical Engineering and Computer Science

• Software Engineering
• Unsupervised Deep Learning
• Artificial Intelligence
• Mechatronic Systems

B.Tech. Mechanical Engineering

National Institute of Technology Hamirpur

Department of Mechanical Engineering

• Engineering Mathematics
• Numerical Methods
• Data Structures

Self-supervised Driver Distraction Detection for Imbalanced Datasets

IEEE 28th International Conference on Intelligent Transportation Systems (ITSC 2025)

Proposed a novel unsupervised adaptation strategy to address class imbalance and enhance generalization in computer vision tasks — directly applicable to clinical AI settings with limited or unlabelled data.

Effect of Size and Cascading of Receivers on the Performance of a Solar Collector System

Springer, Singapore

This study examines how cascading configurations in parabolic dish solar collectors improve collection efficiency by 3.2%–5.4%, highlighting the impact of receiver size and layout on thermal performance.

YOLO-Based Real-Time Object Detection in CARLA

Developed and deployed a real-time object detection system using YOLOv8–YOLOv11 models in CARLA simulator on a Tesla Model 3 setup.

• Simulated a sensor suite with 8 cameras, radar, and 12 ultrasonic sensors at 30 FPS.
• Trained and evaluated 12 YOLO models on a custom dataset; YOLO8-m and YOLO11-m achieved best mAP@(50–95) scores.
• Integrated the fine-tuned YOLO11-m model into the CARLA pipeline for live inference on cars, pedestrians, and traffic signs.

RAG Chatbot using NVIDIA NIM

Implemented a Streamlit-based chatbot leveraging NVIDIA NIM and LangChain for querying research PDFs with contextual awareness.

• Utilized RAPTOR, FAISS, and ChromaDB for retrieval-augmented generation.
• Designed for scalable, domain-specific document understanding.

Codeninja: AI Code Assistant

Developed an LLM-based code assistant using Ollama, Gradio, and LangChain to enhance developer productivity.

• Integrated with a custom model served locally.
• Provided contextual, real-time code generation and fixes.

Implicit 3D Model Generation

Built an implicit 3D object generator using Neural Unsigned Distance Fields (NDF) and latent diffusion.

• Trained on ShapeNet Cars using SLURM on OMNI cluster.
• Implemented vector quantization and feature bottlenecks.

OOD Robustness with AugMix

Evaluated OOD robustness of ConvNeXt and ResNet18 on CIFAR variants using AugMix augmentation.

• Applied transfer learning and tuning with cosine LR scheduling.
• Achieved superior robustness scores for ConvNeXt-tiny.

Real-Time Face Mask Detection

Built a VGG16-based face mask detector with live webcam stream input using OpenCV and Haar cascades.

• Achieved 93.5% accuracy in real-time prediction pipeline.
• Applied transfer learning on VGG16 features.

NSTEDB Research Grant

National Science & Technology Entrepreneurship Development Board (NSTEDB), India

Awarded for the research project 'Effect of Size and Cascading of Receivers on the Performance of a Solar Collector System'. Recognized for innovation in solar energy technology.