Top Repositories
A deep reinforcement learning system for optimizing bridge maintenance decisions across municipal infrastructure fleets, implementing cross-subsidy budget sharing and cooperative multi-agent learning.
Quantile Regression DQN implementation for bridge fleet maintenance optimization using Markov Decision Process. Migrated from C51 distributional RL (v0.8) with 200 quantiles and Huber loss. Features: Dueling architecture, Noisy Networks, PER, N-step learning. All 6 maintenance actions show positive returns with 68-78% VaR improvement.
An experimental platform that structures Japan's River & Sediment Control Technical Standards (Survey / Planning / Design / Maintenance editions) into a Neo4j knowledge graph and compares the performance of GPT-OSS Swallow 20B with and without GraphRAG.
HVAC Range Deviation Forecast - v2.0 (Hybrid Model) Granite TS Embeddings + Statistical Features による高精度異常予測システム
Hybrid Time-Series Anomaly Forecasting Model for HVAC equipment, combining IBM Granite Time-Series Foundation Model (TinyTimeMixer) with statistical feature engineering and temporal attention mechanisms.
建設の技術基準に関する質問の専門性粒度（細かい/粗い）を96%正確に自動判定し、最適なRAGシステム（ColBERT/Naive）を選択する実用的なAgentic RAGシステムのMVPです。2025年11月に公開された河川砂防ダムの技術基準を対象に４つのRAGシステムを構築し、専門性の粒度が異なる200問の質問に対して、精度と速度を比較した。
Repositories
57An experimental platform that structures Japan's River & Sediment Control Technical Standards (Survey / Planning / Design / Maintenance editions) into a Neo4j knowledge graph and compares the performance of GPT-OSS Swallow 20B with and without GraphRAG.
Federated benchmark estimation of bridge deterioration transition probabilities using a Continuous-Time Markov Chain (CTMC) hazard model trained with FedAvg. Raw inspection records never leave each client (municipality). Only 12-dimensional gradient vectors are communicated per round.
This project extends the v2-0 Hybrid Model (statistical features + TTM embeddings) by adding a third modality: text embeddings derived from equipment master data. The three feature vectors — statistical x ∈ ℝ²⁸, TTM embedding y ∈ ℝ⁶⁴, and text embedding z ∈ ℝ¹⁰²⁴ — are concatenated into a 1,116-dimensional triplet feature h and fed into a LightGBM.
HVAC Range Deviation Forecast - v2.0 (Hybrid Model) Granite TS Embeddings + Statistical Features による高精度異常予測システム
Hybrid Time-Series Anomaly Forecasting Model for HVAC equipment, combining IBM Granite Time-Series Foundation Model (TinyTimeMixer) with statistical feature engineering and temporal attention mechanisms.
🏭 Industrial Equipment Classification AI - 78.57% Accuracy with DistilBERT + LoRA
建設の技術基準に関する質問の専門性粒度（細かい/粗い）を96%正確に自動判定し、最適なRAGシステム（ColBERT/Naive）を選択する実用的なAgentic RAGシステムのMVPです。2025年11月に公開された河川砂防ダムの技術基準を対象に４つのRAGシステムを構築し、専門性の粒度が異なる200問の質問に対して、精度と速度を比較した。
Vision Foundation Model for industrial anomaly detection using DINOv2-ViT-Base with LoRA adaptation. Optimized through 4-version experimental study on MVTec-AD2 dataset, achieving stable performance across all 7 categories (average AUC: 0.5626).
Japanese disaster-focused question answering system : Utilizing the bert-base-japanese-v3 + Bi-LSTM + Enhanced Position Heads ultimate architecture, achieving 70.4% End Position accuracy. The combination of Japanese BERT optimization and Bi-LSTM contextual understanding realizes accuracy levels suitable for real disaster response.
A deep reinforcement learning system for optimizing bridge maintenance decisions across municipal infrastructure fleets, implementing cross-subsidy budget sharing and cooperative multi-agent learning.
Quantile Regression DQN implementation for bridge fleet maintenance optimization using Markov Decision Process. Migrated from C51 distributional RL (v0.8) with 200 quantiles and Huber loss. Features: Dueling architecture, Noisy Networks, PER, N-step learning. All 6 maintenance actions show positive returns with 68-78% VaR improvement.
Multimodal RAPTOR for Disaster Documents using ColVBERT & BLIP. Hierarchical retrieval system over 46 tsunami-related PDFs (2378 pages), combining BLIP-based image captioning, ColVBERT embeddings, and GPT-OSS-20b long-context summarization. Optimized for fast multimodal tree construction and disaster knowledge preservation.
This summary outlines the updated professional standards for data scientists in the age of Generative AI, focusing on the shift from technical execution to strategic value creation in Japan, 2026.
This course is designed to teach you how to QUICKLY harness the power of the LangChain library for LLM applications. Build 3 end-to-end working LangChain based generative AI applications with no fluff, no toy examples - just real projects using real APIs and real-world skills.
GPT-OSS B20 Local Execution. Lightweight local environment for running it with Python 3.12 and CUDA acceleration. - Run GPT-OSS B20 entirely offline - Optimize text generation with GPU - Enable fast, secure inference on consumer hardware.
Comprehensive performance analysis of DeepSeek V3 quantization levels (FP16, Q8_0, Q4_0) on 16GB GPU environments.
GPU-accelerated IBM Granite Code model optimization achieving 3-5x performance improvement. Complete benchmarking suite with real-time monitoring and visualization.
A prototype pipeline for generating structured job insights and clustering career trajectories, designed to support agentic AI-based **Proactive Job Goal Creation**.
LLL-based Disaster Detector Agentic AI Application : This project enables the detection and interpretation of environmental threats (e.g., floods, infrastructure risks) by leveraging large language models (LLMs) and multimodal inputs derived from CCTV-based river surveillance feeds.
The MVP provides automated fire risk assessment by extracting wildfire indicators—such as smoke, flame patterns, and thermal anomalies—from imagery, and presenting them in structured natural language analysis.
I've open-sourced Delegator v5.2.1: a state-based AI scheduler designed for large-scale public park maintenance planning in Japan. It converts playground inspection scores into actionable repair/update schedules under budget and workforce constraints — all in under 2 seconds across 1,300+ units. Scalable. Transparent. Real.
This MVP demonstrates a multi-indicator, high-reliability wildfire detection framework that surpasses conventional approaches. By combining Earth observation with intelligent vector analytics, it opens pathways to operational-scale environmental monitoring.
Large-scale fire detection analysis using NASA FIRMS data
Large-scale fire detection analysis using NASA FIRMS data. feat: Add dynamic region support for South America case study in v1-4_area - Enabled flexible geospatial parameterization for wildfire analysis - Updated preprocessing pipeline to accommodate South American satellite data formats.
Large-scale fire detection analysis using NASA FIRMS data feat: Add dynamic region support for EU case study in v1.4.2 - Enabled flexible geospatial parameterization for wildfire analysis - Updated preprocessing pipeline to support EU-specific satellite data formats.
Large-scale fire detection analysis using NASA FIRMS data feat: Add dynamic region support for North America case study in v1.4.3 - Enabled flexible geospatial parameterization for wildfire analysis - Updated preprocessing pipeline to support North America-specific satellite data formats.
Large-scale fire detection analysis using NASA FIRMS data feat: Add dynamic region support for Africa case study in v1.4.4 - Enabled flexible geospatial parameterization for wildfire analysis - Updated preprocessing pipeline to support Africa-specific satellite data formats.
Repair-Aware Survival Analysis: Multi-domain maintenance optimization with NASA CMAPSS & SECOM validation.
EN : Water Pipe Leakage Risk Prediction System v0-3 | Industrial-grade GPR system achieving R²=0.9894 accuracy with 1,907x parallel processing efficiency | Complete scalability proven from N=44 to N=6000 datasets. JP : 水道管漏水リスク予測GPRシステム v0-3 | 産業レベル予測精度R²=0.9894、1,907倍並列処理効率化を実現したガウス過程回帰による高精度漏水リスク分析システム | N=44→N=6000完全スケーラビリティ実証済み
Global Fire Monitoring System v3.2 is an advanced satellite-based fire analysis platform that leverages ESA CEDA Fire_cci data for large-scale global fire pattern detection and clustering analysis. The system processes 12,500+ grid cells simultaneously and provides comprehensive insights into fire behavior patterns across 6 continents.