5+ years turning AI research into products adopted by global industrial leaders
Senior Applied Scientist at Aitomatic
Chat with my AI Twin
I build domain-specific AI systems that solve real problems in industries where general-purpose models fall short. At Aitomatic, I led development of SemiKong and Llamarine, the first open-source LLMs for the semiconductor and maritime industries, now adopted by global industrial leaders including Tokyo Electron and Furuno. My multi-agent framework ProSEA achieved 93.2% accuracy on FinanceBench, outperforming established frameworks like LlamaIndex RAG and LangChain ReAct.
Before Aitomatic, I earned my master's degree from the University of Rochester (advisor: Prof. Chenliang Xu), where my work on object state understanding achieved on par with the commercial SOTA model with a smaller, faster model. I spent nearly 3 years at VinAI Research working under Professor Nguyen Minh Hoai, where I surpassed state-of-the-art on scene text recognition by 3-5% and reduced annotation costs by 50%. I completed my bachelor's with Distinction from Vietnam National University, where I worked with Professor Hoang Van Xiem.
My research spans NLP, computer vision, and machine learning, published at CVPR, NAACL, ACM MM, and AAAI. I currently focus on intelligent AI agents that learn from failure to self-improve, and on specialized LLMs for industries like semiconductors and maritime. SemiKong was featured by VentureBeat, Meta AI Blog, Tom's Hardware, and shared by Yann LeCun.
I'm always happy to discuss ideas and collaborate. Feel free to drop me an email at nguyennm1024@gmail.com.
Domain-specific AI systems that bring expert-level intelligence to industries where general-purpose models fall short.
preprint, 2025
93.2% accuracy on FinanceBench, outperforming LlamaIndex RAG (56.7%), LangChain ReAct (81.6%), and OpenAI Assistants (42.7%). Deployed in production, impressing customers in complex problem solving. Learns from failures to adaptively replan.
JSAI, 2025
The first open-source maritime LLM. Adopted by Furuno for a navigation assistant achieving 100% accuracy on regulation-compliant actions.
OSAI4MU-25 Workshop, AAAI, 2025
The first open-source semiconductor LLM. Adopted by Tokyo Electron for root cause analysis, reducing troubleshooting time by 30%. Featured by VentureBeat and shared by Yann LeCun.
preprint, 2025
Complex real-world problems require more than a single AI call; they demand iterative reasoning. ProSEA is a modular multi-agent framework where a Manager Agent orchestrates domain-specialized Expert Agents, enabling adaptive replanning based on structured feedback from failures and newly discovered constraints. Achieves state-of-the-art performance on the FinanceBench benchmark.
Annual Conference of the Japanese Society for Artificial Intelligence, 2025
The first open-source domain-specific LLM for the maritime industry. Outperforms several commercial products including GPT-4o-mini, GPT-4o, Claude-3.5-Sonnet, and open-source models such as Llama3.1 8B, Llama3.1 70B, and Llama3.3 70B.
OSAI4MU-25 Workshop, AAAI, 2025
The first open-source semiconductor-focused LLM. Outperforms several commercial products including Claude-3.5-Sonnet, Haiku, Opus, Command-R, and open-source models such as Llama3 70B. Featured by VentureBeat, Meta AI Blog, Tom's Hardware, and shared by Yann LeCun.
preprint, 2024
A unified agentic framework that incorporates expert knowledge using neurosymbolic methods to create domain-specific agents with high consistency and accuracy, significantly outperforming ChatGPT assistant and ReAct agent.
ACM MM, 2024
A unified multimodal LLM designed to comprehensively solve tasks related to egocentric video understanding, enabling AI to interpret first-person video from the viewer's perspective.
NAACL, 2024
Introduces a new task for understanding object states and how they change over time. The trained Multimodal-LLM significantly surpasses previous state-of-the-art models and achieves on par with the commercial SOTA model on both GPT-4 and human evaluations.
Arxiv
A simple but effective approach that incorporates language knowledge from large text corpora to improve both text detection and recognition in natural scenes.
CVPR, 2021
A novel approach to incorporate dictionary knowledge in both training and testing phases for scene text recognition. Also introduces VinText, the largest scene text dataset for Vietnamese.
University of Rochester, Department of Computer Science, 2024
Can AI truly understand what happens when you crack an egg or fold a shirt? This thesis tackles a fundamental challenge in vision-language intelligence: teaching models to perceive, describe, and reason about how objects change state. The resulting system achieves on par with the commercial SOTA model using a smaller, faster architecture — pointing toward AI that understands not just what things are, but what happens to them.
US Patent App. 63/726,322, 2024
US Patent App. 63/721,419, 2024
US Patent App. 63/696,337, 2024
WACV 2022, CVPR 2023, CVPR 2024, ACMMM 2024, AAAI 2025, NAACL 2025, CVPR 2025, ACL 2025
VinAI 2021, OSAI4MU AAAI'25
Vietnamese Scene Text Challenge 2021 / Ho Chi Minh city AI Challenge 2021: Vietnamese Scene Text Recognition
Computer vision and machine learning courses for Samsung Display Vietnam's staff, University of Engineering and Technology - Vietnam National University