Haruki Niwa

I work in research and engineering at Hiroshima University, focusing on machine learning, neuroscience, and scientific computing. My work includes developing research software, building computational infrastructure, and creating applications that make research outcomes accessible.

Latest Articles

View all →

Loading articles...

Selected Projects

View all →

Art & Museum Recommender

Personalized artwork recommendations based on individual aesthetic preferences.

Python · NumPy · FastAPI · Jinja2 · JavaScript · Bootstrap · Docker · pytest · Google Cloud Run

StoryChroma

An interactive Japanese audiobook that synchronizes narration, text, and dynamic color visualization of model-estimated emotions.

React · TypeScript · Vite · Python · PyTorch · Hugging Face Transformers · VOICEVOX · Canvas 2D · OKLab

Scene → Art Recommender

Predicting individual artwork preferences from photographic aesthetic judgments.

Python · NumPy · scikit-learn · FastAPI · Jinja2 · JavaScript · Bootstrap · Docker · Google Cloud Run

Flat illustration of two DGX Spark systems connected for distributed DeepSeek inference, with Open WebUI on the worker and an API server on the head.

Distributed LLM Inference on NVIDIA DGX Spark

Deployed DeepSeek-V4.1-Flash-1017 across two NVIDIA DGX Spark systems, with a local API, browser access, and documented PDF evaluation.

NVIDIA DGX Spark · DeepSeek-V4.1-Flash-1017 · TensorFold · RoCE · NFS · Docker · Open WebUI

System architecture showing researchers accessing Open WebUI and Ollama running Gemma 4 31B on a GMKtec EVO-X2 with AMD ROCm GPU acceleration.

Local LLM Inference on AMD Ryzen AI Max

A self-hosted LLM environment built on AMD Ryzen AI Max+ 395 for research use, with similar systems deployed to two laboratories.

AMD Ryzen AI Max+ 395 · Radeon 8060S · Ubuntu · ROCm · Docker · Ollama · Open WebUI · Gemma 4 31B