At Devkraft, I fine-tuned a YOLO-based instance segmentation model on custom datasets spanning eight vehicle damage classes. I also applied LoRA and QLoRA and developed a computer vision pipeline using PyTorch and OpenCV.
I implemented real-time inference and model deployment to automate vehicle damage assessment, reducing manual inspection effort by 80%. I also explored RLHF workflows to align model outputs with human preferences.
As an ML Engineer Intern at Samsung R&D, I developed a multimodal face-voice model for cross-lingual speaker recognition that achieved 28.46% EER and ranked in the top four. A two-stream architecture with gated fusion reduced model variance by 18%.
At Samsung ResearchLab, I built LW-DMNet, a lightweight CNN for dementia classification that achieved 98.73% accuracy with 90% fewer parameters. The work was published in Springer Lecture Notes in Networks and Systems, and received the Best Paper Award at the International Conference on Applied Artificial Intelligence (2AI2026).
