About Me

FUJITAKE Masato (藤武将人) is a Co-Founder of Fast Accounting, Ltd., Tokyo, Japan, and Research Director of its research division, FA Research.

His research focuses on the end-to-end development of high-performance domain-specific large language models (LLMs), particularly for accounting and financial applications. His work spans the entire model development pipeline, including data collection and corpus construction, pretraining from scratch, mid-training, post-training, reinforcement learning, and context-window extension.

He has been developing Deep Dean, an accounting-specialized LLM that has substantially outperformed general-purpose frontier models in selected accounting-domain evaluations.

To advance domain-specific LLMs, he also investigates generalizable methods for improving model performance, training efficiency, and stability. His recent research includes EnergyCal and MESH, focusing on optimization techniques for Mixture-of-Experts (MoE) training. He also develops and releases experimental general-purpose language models through Hugging Face. While the domain-specific models themselves are proprietary, he actively publishes broadly applicable methods and research findings derived from their development.

Previously, his research centered on computer vision, optical character recognition, and document understanding, resulting in publications at international conferences including WACV, ICASSP, and LREC-COLING.

He received his Ph.D. in Information Science and Technology from the Graduate University for Advanced Studies (SOKENDAI) in 2022, where he conducted research at the Computer Vision and Discrete Geometry Group (Sugimoto Laboratory) at the National Institute of Informatics (NII). He earned his bachelor’s and master’s degrees from Shibaura Institute of Technology in Tokyo, Japan.

He has also contributed to open-source projects, including torchstat and Nobunaga.

Research Interests

Domain-Specific Large Language Models
End-to-End LLM Development and Training
Data Curation and Corpus Construction
Pretraining, Mid-Training, and Post-Training
Reinforcement Learning and Long-Context Modeling
Efficient Optimization and Mixture-of-Experts (MoE)

News (Selected Updates)

2026/10: My paper “EnergyCal: Taming Loss Spikes in Muon-Family Optimizers for Sparse MoE Training” is available on arXiv.
2026/08: My paper “MESH: Memory-Efficient Sinkhorn Optimization for Mixture-of-Experts Training” is available on arXiv.
2026/07: Our experimental MoE language model, Rize-0.6-tiny, is publicly available on Hugging Face.
2026/03: Our earlier research model, Rize-0.5-tiny, is available on Hugging Face.
2026/01: Our accounting-specialized LLM, Deep Dean, achieved 99.8% accuracy on the Japanese Bookkeeping Level 1 examination. Details
2025/04: We announced Deep Dean, an accounting-specialized LLM demonstrating strong performance on professional accounting examinations, including the USCPA. Details
2024/09: My paper “JaPOC: Japanese Post-OCR Correction Benchmark using Vouchers” was accepted to PRICAI 2024. arXiv
2024/04: My paper “JSTR: Judgment Improves Scene Text Recognition” was published on arXiv.
2024/03: My paper “LayoutLLM: Large Language Model Instruction Tuning for Visually Rich Document Understanding” was accepted to LREC-COLING 2024. arXiv

For a complete list of publications, see my arXiv author page.

Education

Ph.D. in Information Science and Technology, 2022: The Graduate University for Advanced Studies, SOKENDAI
Master of Electrical, Electronic and Communications Engineering, 2019: Shibaura Institute of Technology, Tokyo
Bachelor of Electrical Engineering Technologies, 2017: Shibaura Institute of Technology, Tokyo

Work Experience

Research Director, 2024/01 ~ present, Fast Accounting, Japan
Chief Research Scientist, 2023/02 ~ 2024/01, Fast Accounting, Japan
Research Scientist, 2022/04 ~ 2023/02, Fast Accounting, Japan
Research Assistant, 2020/04 ~ 2022/02, National Institute of Informatics, Japan
Research Engineer, 2016/11 ~ 2022/03, Fast Accounting, Japan