Teppei Nakano

Resume

As of February 15, 2025
Name: Teppei Nakano

■ Career Overview
After graduating from university, I entered graduate school to research machine learning model development and applications in healthcare data. During my studies, I submitted and had papers accepted at NeurIPS and ICML workshops. Over approximately five years, including graduate school, I worked on five projects mainly focusing on system development and operation.

■ Strengths
- Creating and implementing data pipelines for machine learning models
- Defining requirements, designing, developing, and supporting the introduction of machine learning systems
- Project management (Experience managing projects with up to 10 members)

Company: DREAM FINANCE Inc. (June 2024 – January 2025)
Business: Development of financial systems Capital: 1 billion yen | Revenue: 20 billion yen (2024) | Employees: 50
Project (June 2024 – January 2025, 8 months): Blockchain Analysis Business Using Machine Learning in Hong Kong
Developed a blockchain data analysis platform using machine learning to monitor cryptocurrency transactions.
Team: 85 person-months / 9 in-house members / 3 vendors
Responsibilities:
Designed and developed an anomaly detection system for blockchain using machine learning.
Built algorithms to analyze transaction data in real-time and identify suspicious activities.
Developed tools for visualizing relationships between wallet addresses using network analysis.
Constructed a system for risk scoring based on transaction history data.
Implemented an AI model to automatically detect security risks in smart contract codes.
Promoted data-sharing projects to improve transparency in the industry by partnering with local blockchain firms.
Achievements:
Improved detection accuracy of suspicious transactions by 35%, enhancing AML compliance.
Detected large-scale fraudulent networks through wallet address analysis and collaborated with law enforcement.
Automated detection of fraudulent market transactions and streamlined smart contract security assessments.
Proposed data-driven risk management strategies adapted to regulatory changes in Hong Kong’s crypto industry.
Company: Cahall Inc. (April 2021 – May 2024)
Business: Development and delivery of machine learning models Capital: 10 million yen | Revenue: 200 million yen (December 2023) | Employees: 10
Project (January 2023 – May 2024, 14 months): Development of Privacy-Focused Local LLM
Designed and developed a local LLM for insurance and healthcare companies, ensuring high confidentiality.
Team: 60 person-months / 6 in-house members / 2 vendors
Responsibilities:
Evaluated major local LLMs (Llama by Meta, Gemma by Google, Qwen by Alibaba) and selected Llama-3-8B-Instruct.
Built a custom model pipeline using PyTorch and fine-tuned with LoRA (Low-Rank Adaptation) via Hugging Face tools.
Collected and curated a specialized dataset for medical terminology to fine-tune the model.
Optimized inference for resource-constrained devices using quantization techniques with Llama.cpp.
Resolved tokenizer inconsistencies and library update issues for a stable environment.
Developed medical data summarization and Q&A features, achieving high accuracy in benchmark tests.
Achievements:
Implemented an automatic summarization feature for medical documents, reducing review time for specialists.
Developed a system for analyzing medical data while preserving privacy, successfully adopted in hospitals.
Enabled on-premises operation of the LLM, eliminating the need for external data transmission.
Conducted real-world tests in medical environments, improving the system based on user feedback.
Project (January 2022 – December 2023, 12 months): AI-Driven Personalized Marketing Platform
Developed a system to enhance marketing strategies through AI-based consumer behavior analysis and personalized ad delivery.
Team: 70 person-months / 7 in-house members / 3 vendors
Responsibilities:
Developed an AI engine to analyze consumer data in real-time and deliver optimized ads.
Designed and implemented a customer segmentation model using machine learning.
Created a social media analysis tool using NLP to automatically extract trends and customer preferences.
Built an algorithm to automate A/B testing, speeding up marketing effectiveness evaluations.
Utilized edge AI to accelerate ad delivery and optimize costs.
Developed a corporate dashboard to visualize ad performance and provide real-time reports.
Ensured GDPR and CCPA compliance with a privacy-conscious design.
Achievements:
Improved click-through rates (CTR) by 25%.
Increased conversion rates (CVR) by over 30% through personalized ads.
Reduced marketing analysts' data analysis time by 40% with the corporate dashboard.
Enhanced social media trend analysis accuracy, accelerating consumer insight acquisition.
Project (January 2021 – December 2021, 12 months): Legal Document Search System Using Large-Scale Transformer Models
Developed an AI system to search, summarize, and analyze legal documents, streamlining legal and compliance operations.
Team: 75 person-months / 7 in-house members / 2 vendors
Responsibilities:
Designed and implemented a legal document search algorithm using large-scale transformer models (BERT, GPT-3).
Prepared datasets of laws, precedents, contracts, and regulations, and built a legal-specific BERT model.
Developed a question-answering search system using NLP.
Created an embedding model for similar case searches, improving legal research accuracy.
Developed a dashboard for legal professionals with ranking and summarization features.
Ensured data privacy and security for on-premises operation.
Achievements:
Improved legal document search accuracy by 20%.
Reduced legal document reading time by over 50% with AI summarization.
Enhanced search efficiency by 30% using transformer models.
Reduced legal research time and improved compliance efficiency through similar case searches.
Automated risk analysis in contract reviews, identifying potential legal risks.
Position: Distinguished Researcher, Graduate School of Medicine, University of Tokyo (April 2018 – March 2021)
Project (April 2018 – March 2021, 36 months): Development of Automated Medical Record Summarization AI
Conducted research on summarizing electronic medical records using machine learning to improve information organization in medical settings.
Responsibilities:
Designed and implemented an automated summarization model for electronic medical records using BERT.
Pre-trained models with self-supervised learning specialized in the medical field.
Fine-tuned Seq2Seq models based on Transformer architecture.
Collected and anonymized medical record data for training datasets.
Developed a hybrid approach combining rule-based systems with machine learning for accurate medical term summarization.
Collaborated with hospitals to evaluate real-world applicability.
Optimized models based on metrics like ROUGE scores.
Achievements:
Reduced physicians’ documentation organization time by 40%.
Improved ROUGE scores by 15% and validated high accuracy through expert reviews.
Developed a user interface allowing customization of summary results.
Achieved real-world deployment readiness through joint research with hospitals.
Reduced the burden of data labeling by applying self-supervised learning while improving domain adaptation.
■ Certifications
Regular Driver's License (MT)
TOEIC Score: 950 (June 2017)
TOEFL iBT Score: 108 (June 2017)
■ Education
Bachelor of Medicine, Keio University (March 2017)
Graduate School of Medicine, University of Tokyo (Withdrew in March 2021)
■ Publications
T. Nakano, K. Ikeda, “Predicting Cancer Heterogeneity from One-shot Biopsy,” ICML Workshop on Computational Biology, 2017.
T. Nakano, K. Ikeda, “Visualizing Cancer Heterogeneity with Dynamic Flow,” ICML Workshop on Computational Biology, 2017.
R. Sato, T. Nakano, et al., “RNA Sequencing Analysis Reveals Interactions between Breast Cancer or Melanoma Cells and the Tissue Microenvironment during Brain Metastasis,” BioMed Research International, 2016.
T. Nakano, “Generating Clinical Texts from Conversation,” NIPS Workshop on Machine Learning for Health, 2016.
M. Hashimoto, T. Nakano, H. Yashiro, T. Fujita, and M. Jinzaki, "Application of Word2vec to Radiology Report Analysis," Proceedings of the Japan Association for Medical Informatics Annual Conference, vol. 35, pp. 904-905, November 2015.
T. Nakano, “Early Detection of Hepatorenal Syndrome from Medical Records,” NIPS Clinical Workshop, 2015.
T. Nakano, “Unsupervised Feature Detection for Medical Images,” 28th Annual Conference of the Japanese Society for Artificial Intelligence, 2014.
■ Skills and Competencies Problem-Solving:
Experience in end-to-end system development using machine learning and AI, including data preprocessing, model selection, accuracy improvement, and deployment.
Special focus on NLP and image recognition using large-scale transformer models, improving accuracy and optimizing computational costs.
Successfully improved summarization accuracy by 30% in medical data analysis projects using self-supervised learning.
■ Coordination:
Strong coordination skills for managing diverse stakeholders such as engineers, data scientists, and domain experts.
Built consensus on data quality and AI output interpretation to ensure practical system development.
Emphasized collaboration with legal experts to ensure the reliability of legal document search AI.
■ Persistence:
Applied iterative approaches like hyperparameter tuning, data preprocessing, model ensembling, and resource optimization to enhance model performance.
Improved document search accuracy by 20% by refining data cleaning and fine-tuning processes in transformer-based models.
■ Personal Statement
I actively seek to learn new technologies and apply them in practice. In machine learning, where techniques evolve rapidly, I’ve honed skills to quickly investigate and implement optimal methods. I share knowledge within teams to support swift decision-making.
I value smooth communication with stakeholders, clearly explaining the rationale behind my proposals and ensuring mutual understanding. I strive to present complex technical content in an accessible manner to maximize project outcomes.
When facing challenges, I focus on identifying the core problem, formulating hypotheses, and deriving effective solutions. For instance, when machine learning models underperform, I revisit data preprocessing and explore alternative approaches to efficiently resolve issues. I aim to continue applying my experience to tackle challenges flexibly and effectively.