cv

AI Innovation Lab Tokyo, First Cohort (Contract)

Supercell

Related: Generative AI, Game Development, Python, TypeScript, Next.js, Firebase, GCP

• Planning and developing a new AI-centered game with technical and financial support from Supercell.
• Leading the design and validation of LLM and generative-AI features in a two-person team, and driving prototype development.

Machine Learning Engineer Intern

MinD in a Device

Related: Python, PyTorch, Docker, AWS, Weights and Biases, 3D Reconstruction, SLAM

• Evaluating and improving 3D reconstruction and SLAM models on customer data under real deployment conditions, while maintaining experiment infrastructure.
• Proposed and implemented a catenary-curve fitting method based on the physical shape of power lines to address unstable 3D reconstruction from sparse power-line point clouds; the method was adopted in the production system.
• For an underwater-structure reconstruction project, proposed using COLMAP Underwater, a COLMAP variant specialized for underwater scenes, and improved reconstruction quality by tuning settings and parameters for domain-specific constraints.
• Consolidated official implementations of multiple 3D reconstruction and SLAM models into one repository and built a reproducible Docker-based benchmark environment; compared GPU memory, runtime, and accuracy under the same conditions to recommend models for each project.
• Built reusable experiment infrastructure with Weights and Biases and improved comparison and sharing of evaluation results through visualizations such as segmentation-result overlays.

Software Engineer Intern

freee

Related: TypeScript, Next.js, React, Ruby on Rails, AWS, Git

• Worked on the IAM (Identity and Access Management) team, developing authentication and authorization infrastructure used across all products.
• Designed an authentication flow using passkeys to improve the login experience and prototyped passkey login in a test environment.
• Documented user scenarios and adoption concerns around passkeys, including device loss and registration flows.
• Created and merged a pull request improving the internal design system, increasing UI consistency and reusability.

Application Engineer Intern

Rakuten Group

Related: HTML, CSS, JavaScript, Django

• Planned, designed, and implemented a new web application with a team in a short-term hackathon-style internship.
• Worked end-to-end from requirements definition to prototype development and implementation, mainly owning backend logic.

M.S. in Information and Communication Engineering

The University of Tokyo

Graduate School of Information Science and Technology, Department of Information and Communication Engineering, Echizen Lab

• Conducting research on AI safety as a master's student and research assistant.
• Researching image protection methods against misuse of image-to-video models, such as sexual deepfakes and copyrighted-content generation, by adding imperceptible noise to input images to disrupt motion consistency in generated videos.
• Designing, implementing, and validating evaluations to balance generation quality and defensive effectiveness.
• Related: Research, Python, Generative AI, AI Safety

B.S. in Information and Communication Engineering

The University of Tokyo

Faculty of Engineering, Department of Information and Communication Engineering, Matsui Lab

• GPA: 3.86 / 4.30
• Conducted undergraduate research on pruning techniques for decision-tree ensemble models.
• Proposed a pruning method that achieves a strong trade-off between prediction agreement before and after pruning and compression rate; the work was accepted to ICML 2026 (CORE: A*).
• 6,352 / 23,918 submissions were accepted, for an acceptance rate of 26.6%.
• Related: Research, Python, Optimization, Machine Learning, Writing

English Language Study Abroad Program

The University of Sheffield

English Language Study Abroad Program

• Completed a one-month English language program and attended English classes.
• Built communication skills and cross-cultural understanding through interaction with local students.

Gemini 3 Tokyo Hackathon | 3rd Place (out of 70 teams)

Google DeepMind, Supercell, and Cerebral Valley

Developed PrompDojo, a real-time competitive prompt quiz game, at a hackathon for building new products with Gemini.

PINE: Pruning Boosted Tree Ensembles with Conformal In-Distribution Prediction Equivalence

Meeting on Image Recognition and Understanding (MIRU)

Haruki Yajima and Yusuke Matsui. Domestic conference, peer-reviewed, under review.

PINE: Pruning Boosted Tree Ensembles with Conformal In-Distribution Prediction Equivalence

International Conference on Machine Learning (ICML)

Haruki Yajima and Yusuke Matsui. International conference, peer-reviewed.

PrompDojo | Gemini 3 Tokyo Hackathon

Related: Next.js, TypeScript, Firebase, GCP, Gemini API, Git

• Developed a real-time competitive prompt quiz using Gemini and won 3rd place out of 70 teams.
• Designed a new AI x Education x Game experience where players learn prompt engineering through competition.
• Led specification design, technology selection, and implementation based on the team's ideas.
• Links: Hackathon Gallery: https://cerebralvalley.ai/e/gemini-3-tokyo-hackathon/hackathon/gallery | Project Page: https://haruk1y.github.io/prompdojo-project-page/

Kotone's Atelier | Mistral Worldwide Hackathon - Tokyo edition

Related: Next.js, TypeScript, ElevenLabs API, Mistral Fine-Tuning, Hugging Face, Weights and Biases

• Developed Kotone's Atelier, a game where players request music from abstract prompts, with the goal of making music generation more approachable.
• Gamified the music-generation experience by combining music generation through the ElevenLabs API with request interpretation and in-game evaluation using a fine-tuned small Mistral model.
• Owned ideation, specification writing, technology selection, and implementation, designing the game loop and AI pipeline end-to-end.
• Links: Hackathon Page: https://luma.com/mistralhack-tokyo?locale=ja | GitHub: https://github.com/Haruk1y/Mistral_Hackathon