$ whoami
AI Architect / Engineer · Medical AI Researcher · CTO / Board Member
As an AI architect/engineer and AI development lead, I design and build generative-AI products and computer-vision systems, while researching medical AI (mammography, EEG, and LLMs on electronic health records) at Tohoku University Graduate School of Medicine and RIKEN AIP. As a startup CTO and board member, I have owned everything from requirements and design to implementation and operation.
Ph.D. Student, Tohoku University Graduate School of Medicine · Statistical Genetics Team, RIKEN AIP · Board Member, BINARISTA, Inc.
PdM / AI & web development (team of 5)
Built and operate an enterprise sales-support AI-agent web application from scratch, combining product management with development, and acquired new customers.
BINARISTA, Inc. (Board Member)
Next.js / FastAPI / LangChain / LangSmith / Gemini API / PostgreSQL
AI development lead (team of 5)
Developed several generative-AI features, including LLM-based chat and document revision, and shipped them into a service used by 100k+ users.
HR service company (contract)
Next.js / FastAPI / GCP / Vertex AI / Gemini / MySQL
AI development (team of 5)
Built an NLP-based AI that extracts multiple fields from invoice PDFs, improving extraction accuracy from 40% to 80%.
Climate-tech SaaS company (contract)
Python / PyTorch / spaCy / GCP
AI development lead (team of 8)
Owned an image-processing AI (person recognition and editing) from development planning to implementation, improving person-detection accuracy from 90% to 96%.
Entertainment company (contract)
Python / PyTorch / TensorFlow / GCE
AI development lead (team of 10)
Planned and built an end-to-end AI pipeline, improving accuracy from 65% to 90% and reaching on-site deployment. Also launched the data-science team.
NTT DATA Corporation (full-time)
PyTorch / TensorFlow / Databricks / React / Flask
AI development (team of 5–7)
Developed anomaly detection for manufacturing image and time-series data. EDA-driven improvements reached 85%+ accuracy in automotive-part defect detection.
NABLAS Inc. (full-time)
Python / PyTorch / TensorFlow / scikit-learn / tslearn
I research medical AI at Tohoku University Graduate School of Medicine. My main topics are breast cancer detection in mammography, differentiating epileptic seizures with EEG, and applying LLMs to electronic health records. I publish peer-reviewed papers through joint research with physicians.
ArticleFirst author
Y. Kashiwada, E. Takaya, M. Hiroya, et al. — PeerJ Computer Science, 11:e3252
MIME-ViT fuses four Vision Transformer branches with different patch sizes and a CNN branch to localize masses and calcifications on mammograms. Trained from scratch on the TOMPEI-CMMD annotations, it reached a mean IoU of 0.3342: better than DETR without pre-training, but below pre-trained DETR and YOLOv8x.
DatasetFirst creator
Y. Kashiwada, E. Takaya, M. Hiroya, et al. — The Cancer Imaging Archive (TCIA)
TOMPEI-CMMD adds polygon lesion segmentations (mass, calcification, and four other finding types), lesion locations, corrected laterality/view labels, and clinical data to the Chinese Mammography Database (CMMD). It contains 1,773 masks for 1,385 breasts from 1,363 patients and is published on TCIA under CC BY 4.0.
ArticleSecond author (co-author)
K. Konomatsu, Y. Kashiwada, T. Kubota, et al. — Epilepsy & Behavior, 172:110713
Ictal EEG was turned into time-frequency images (density spectral arrays, DSA) and a ResNet34 model was trained to tell mesial temporal lobe epilepsy (mTLE) seizures from functional/dissociative seizures (FDS). Among reduced-electrode settings, the single Cz channel gave the highest AUC, 0.941.
Read the summary: Single-channel EEG DSA for seizure differentiation →DOI: 10.1016/j.yebeh.2025.110713 ↗
→ The full list of publications, including talks, awards, and patents, is in my CV.
All research →Design, implementation, and production operation of AI agents, RAG, and LLM-powered features
Model development and accuracy improvement for computer vision, NLP, and time-series analysis
AI adoption strategy, technology selection, AI security, and building development teams
Talks on generative AI and medical AI; joint research with universities and companies
Feel free to reach out about technical consulting, AI development, advisory roles, joint research, or speaking requests. I usually reply within 2–3 business days.