top of page

AI-Powered Management Decision Support System (MDSS) Development Notes

Jun 24
3 min read

I am validating whether it is possible to build a system that supports management decisions, rather than just a chatbot, by incorporating AI APIs into business operations. The purpose is to provide materials for decision-making by integrating information such as sales, markets, competitors, and news. It is assumed that the final decision will be made by a human.

  • Current Status Analysis

  • Risk Analysis

  • Fishbone Analysis

  • Mind Map Generation

  • PDCA Proposal


System Architecture 

The current PoC configuration is as follows, though in a corporate environment, the basic data would likely be managed via Power BI or an internal database instead of SQLite: SQLite (or Power BI / Internal DB for corporate use), Python, Gemini API, Streamlit UI


Implementation Details

Database construction (SQLite): Created the following tables for test data:

  • sales_transaction

  • customer_master

  • product_master

  • region_master

  • stock_prices

  • economic_news

  • competitor_status


AI Analysis Features 

Built a mechanism to generate the following tools based on sales data (sales_transaction)

  • Weak Point

  • PDCA

  • Mindmap

  • Fishbone


UI Construction 

Implemented the following screens using Streamlit:

  • Sales

  • Market / Macro

  • Competitor

  • PDCA

  • Mindmap

  • Fishbone


Challenges Encountered During Development

1. Gemini API Rate Limit

Occurred 429 RESOURCE_EXHAUSTED!!! Initially, API calls were made at every opportunity: Screen display > API call, Tab switching > API call, Question > API call This is because I am using Streamlit for the browser UI, and Streamlit re-executes the script upon every interaction, causing unintended Gemini API calls.


2. Cost and Operational Efficiency 

Although I aimed for real-time analysis, it became clear that a design calling the API every time resulted in the following issues:

  • Increased costs

  • Rate limiting

  • Reduced response speed


Improvement Strategy

Improvement Strategy Shifted from AI-centered design to Database-centered design

  • Before: User action -> (API call) -> Real-time analysis (High cost, unstable)

  • After: User action -> (DB reference) -> Display analysis results + Use AI only upon updates (Cost optimization, stable operation)


Minimize AI Usage

AI is used only for update requests from the user, with results saved to the database. Daily references to the following are retrieved from the database. In addition, the design was revised to provide faster responses by using caching.。

  • Analysis Generation

  • Summary

  • Fishbone creation

  • Mind map creation

  • PDCA generation


Future Vision

Ultimately, I aim to create an interactive decision-support system that integrates sales, customer, product, news, stock price, exchange rate, and competitor information, providing analysis and advice in the format of "In my opinion."


The key is that the AI does not make the final judgment; rather, it organizes information and generates hypotheses for humans to make the decision.

 

Summary

Having actually built it, I realized several things. Even if an idea is good, it may be rejected when proposed to clients or the company. However, there are reasons for this. Decisions are made from various perspectives, not just technical feasibility:

  • Is it operational?

  • Is it maintainable?

  • Is the cost reasonable?

  • Will it truly lead to profit?


During the development of the MDSS, I encountered API limits multiple times.

Using AI > Calling the API excessively > Reaching usage limits > Being prompted to upgrade to a paid plan.


That led me to consider, "Is it worth paying that cost?" This perspective applies to both companies and individuals. Ultimately, a solution is adopted only when the benefits gained from its implementation outweigh the costs; if the effects are unclear, it is natural for it to be put on hold or rejected. I thought I understood this intellectually, but by building it myself, I was able to truly realize it.


Learning to consider perspectives beyond just technology—such as operations and return on investment (ROI)—through the development of AIOCR and MDSS has been a significant takeaway.


Furthermore, I am proceeding with development by handling the design myself, having the AI write the code, and then performing tests and reviews. I believe this way of working will likely become more prevalent in the future.

While it is desirable to be able to write code, I believe it is even more important to be able to understand "what the code is doing" and "what is missing."


Without that understanding, no matter how much code the AI writes, you will never be able to reach completion.。


Through this development, I was able to reconfirm the importance of not only technology but also design, operations, costs, and reviews. I believe that this alone made the challenge well worth it.

Recent Posts

See All

Comments


bottom of page