At Indian Institute of Technology Ropar, I conducted research for a survey on Group Affect Recognition, reviewing 100+ studies, 12+ state-of-the-art methods, and 15+ benchmark datasets. I contributed to a structured taxonomy and future research roadmap for the field in computer vision.
On my Mistral-7B project, I fine-tuned an open-weight model for mathematical reasoning using an SFT-to-GRPO pipeline on GSM8K. It achieved 52.5% test accuracy, an 11% improvement over the zero-shot baseline.
I also designed Mirage, a multimodal deepfake video detection pipeline that combines audio and video with cross-attention. It achieved 75% detection accuracy on GAN-based and diffusion-based deepfakes.
For SentinelVision, I designed a real-time PPE compliance pipeline that turns video detections into structured violation records. Its confidence fusion and hysteresis thresholds reduced false positives by 60%.
