Applied AI Research Engineer; vision-language models, computer vision, efficient deep learning.

I work independently on vision-language models: building them from scratch and testing design choices with controlled experiments. Computer vision is my base; model efficiency (knowledge distillation, parameter-efficient fine-tuning) is the background I bring to it.

Background in Electrical & Electronics Engineering.

MicroCLIP — from-scratch CLIP-style VLM Image and text encoders trained from scratch on COCO Captions, with an 11-configuration sweep comparing sigmoid (SigLIP) and softmax contrastive losses under a fixed compute budget. Small-scale non-replication: softmax beat sigmoid at every batch size, and the gap widened as batch size shrank. Live demo · checkpoints on Hugging Face
PyTorch · CLIP · SigLIP · Contrastive Learning · ONNX

RT-DETR Knowledge Distillation — KD for real-time detection Logit-, feature-, and combined-distillation strategies for RT-DETR, with reproducible experiments and deployment-oriented evaluation. First-author preprint: arXiv:2605.31191
Knowledge Distillation · RT-DETR · Object Detection · PyTorch

detrflow — training-to-serving stack for RT-DETR Training, evaluation, benchmarking, and deployment with FastAPI/Gradio serving and Docker.
Computer Vision · FastAPI · Docker · PyTorch

• Hugging Face PEFT — #3293 · Merged — removed redundant CUDA memory overhead in prepare_model_for_kbit_training (~480 MB lower memory, measured independently by the issue reporter)
• aisuite — #319 · Merged — Python 3.14 compatibility