top of page


AI-Powered Management Decision Support System (MDSS) Development Notes3
今日は、AIを活用したManagement Decision Support System(MDSS)のチャット機能を改善しました。これまでの課題は、「Hello」のような挨拶でもAIが経営分析を始めてしまうことでした。原因は、通常チャットでも分析前提のプロンプトを使用していたためです。今回、業務に関する質問のみ分析データを参照し、それ以外は通常の会話として応答するよう設計を見直しました。一方で、What-if分析やフォローアップはLLMによる推論が不可欠なため、APIを呼び出す構成としています。AI・データベース・アプリケーションの役割分担を設計することが、使いやすさと運用コストの両立につながります。
Jul 32 min read


AI-Powered Management Decision Support System (MDSS) Development Notes2
During this development cycle, I solved a Gemini API 429 error, redesigned the MDSS architecture, and fixed a Streamlit layout issue. The API problem was caused by using the wrong model (gemini-2.0-flash instead of gemini-2.5-flash). After correcting it, the application worked normally. I also changed the system from AI-first processing to a database-first design, using AI only for updates to reduce API usage and improve response time. Finally, I removed duplicate chat render
Jun 282 min read


AI-Powered Management Decision Support System (MDSS) Development Notes
This project aims to build a management decision support system by integrating AI APIs into business operations. By synthesizing data on sales, markets, competitors, and news, the system automates the structuring of complex management issues and hypothesis generation through tools like status analysis, PDCA proposals, and mind mapping, providing high-quality materials for human decision-making.
During development, I addressed practical challenges such as API limits and costs
Jun 243 min read


AIOCR
2年越しにAI-OCRの構築へ着手しました。今回は圧倒的な学習量を誇る大手AI(Gemini-2.5-flash)を使い、データ構造をガチガチに固定する設計です。
核となるのは「判定 ➔ 抽出」の2段階パイプライン。1回目で帳票種別と顧客を判定し、2回目で顧客ごとのルールを動的に適用して高精度に抽出する構想です。
手始めに、手元のPDFから「未知の顧客(UNKNOWN)」として項目を抜き出すベース処理をテストしたところ、狙い通り綺麗なJSONが出力されました!
Jun 142 min read


IF YOU CAN WRITE THE FLOW YOU CAN DO ANYTHING
- Realizing my former boss's words in PowerShell I created a simple analysis script in PowerShell to inspect the contents of data involving character encodings. The motivation was a situation where I needed to verify file contents in a human-readable format, and I took the opportunity to try it out as part of my learning. While the implementation itself was partially generated by AI, I provided the overall flow (the comments in the PowerShell file at the bottom) and the encod
May 173 min read
bottom of page