A computer‑vision and deep‑learning engineer working at the seam between research and production — currently at Eurofinsin Ho Chi Minh City, recently a student researcher in Tokyo, presently chasing a Master's in AI.
My name is Nguyễn Văn Lộc — a professional and enthusiastic programmer with a particular love for Artificial Intelligence, especially Computer Vision and Deep Learning. I am a quick learner with a self-learning attitude, always eager to explore new technologies and passionate about solving hard problems. Python is my core language and closest tool; I love almost everything that can be built with it.
Over the last several years I've moved between a research lab at the University of Science in Ho Chi Minh City, a student researcher position at the National Institute of Informatics in Tokyo (Echizen Laboratory, under Prof. Isao Echizen), a role building AI agents at Autonomous, and now a Software Engineer position at Eurofins. The throughline is the same: take a model out of the notebook and into something a person can use without thinking about it.
I read widely, ship steadily, and believe that the most interesting problems sit between disciplines — vision and language, research and product, Vietnamese and Japanese, theory and the messy thing it's supposed to model.
Addressing the Ho Chi Minh City AI Challenge 2025, Visionary introduces an adaptive keyframe extraction algorithm, a Qwen3‑VL pre‑processing pipeline for metadata generation with integrated OCR, a flexible architecture supporting multiple embedding models, and Reciprocal Rank Fusion for synthesizing retrieval results across complex, large‑scale video tasks.
A practical pipeline for classifying ENT endoscopy imagery that pairs an ensemble of vision backbones with a curated augmentation schedule, designed to remain robust to the imaging variance typical of clinical capture devices.
A zero‑shot object removal method that leverages pixel‑wise attention dissolution and latent guidance within a diffusion framework — achieving clean inpainting without task‑specific fine‑tuning on the target scene.
A demonstration of a single‑pass removal approach for visually similar object clusters in natural imagery, extending zero‑shot inpainting toward the messy, repetitive scenes that one‑object methods choke on.
Working on something that needs computer vision, agent infrastructure, or a careful pair of hands? Let's talk — research collaborations, contract work, or just an interesting question.