DEV Community

Leonhard Satria Suharjo
Leonhard Satria Suharjo

Posted on

How I Connected AI Agents with Manufacturing Supply Chain

#googleaichallenge #ai #python #kaggle

Building SuplaiCerdas: From Google-Kaggle’s AI Agents Course to Real-World Supply Chain Automation

*Project view on Github @leonhardsuharjo

Weeks ago, I had the opportunity to participate in the AI Agents 5-day Intensive Course held by Kaggle and Google. The course consisted of daily notebook exercises, video podcasts, and livestreams that systematically introduced me to the fundamentals of AI agents — from basic concepts to advanced topics like agent tools, multi-agent architectures, and quality evaluation.

This article serves as a part of my capstone project submission: SuplaiCerdas, a prototype AI agent application designed to automate material cost and risk assessment for manufacturing projects.

Further details can be found at my github profile @leonhardsuharjo

My Project in 34 Words

An enterprise AI agent system that monitors component prices across suppliers for IoT manufacturing companies. The system helps maintain profit margins on tender commitments by detecting price spikes before they impact production costs.

Background issue — The Manufacturing Business Challenge

As someone deeply interested in the device and manufacturing business model, I’ve observed a critical challenge that many companies face during project execution.

Manufacturing projects often involve months-long processes with multiple phases: tender submission, designing, testing, evaluation, and mass production. Here’s the catch — companies must commit to specific prices during the tender phase, which happens months before actual project execution.

The risk
Suppose a scenario where a business submits a tender for smart TV controllers at a price of 5000 USD each and win the bid. However, months later when production actually started, the price of the smart TV’s critical component, suppose its microcontroller has spiked twice in price due to various reasons. Therefore, this will bring disruption to the company’s initial budget that was submitted during the bid, resulting in margin loss.

SuplaiCerdas is created to address this risk.

*Sample material price

What I Learned (and Applied) from the Course

*Course main page

Before diving into the project, I completed three key notebooks from the 5-day course that directly shaped my approach:

Day 1B: Multi-Agent Systems & Workflow Patterns

*Kaggle notebook accessible at my profile @leonhardsuharjo

This notebook introduced me to multi-agent architectures — the idea of building specialized agents that work together like a team. The key concept I applied from this notebook is Sequential Workflow Pattern, which is breaking down tasks into a pipeline of different agents performing sequential tasks.

For SuplaiCerdas, this meant:

  1. PriceCheckAgent retrieves current prices from all suppliers
  2. ComparisonAgent calculates price changes and assigns risk levels
  3. RecommendationAgent synthesizes findings into actionable procurement advice

Day 2A: Building Custom Function Tools

*Kaggle notebook accessible at my profile @leonhardsuharjo

This notebook taught me that agents become powerful when they can use tools to interact with the outside world. The key concept I applied in this notebook is Custom Function Tools, which is turning python functions into agent tool using ADK (Agent Development Kit).

For SuplaiCerdas, I built two custom tools:

  1. check_component_prices(): Connects to supplier database and retrieves current prices
  2. calculate_price_change(): Calculates percentage change from historical data and flags risk levels (HIGH >10%, MEDIUM 5–10%, LOW <5%)

Day 4B: Agent Evaluation

*Kaggle notebook accessible at my profile @leonhardsuharjo

This notebook emphasized that agents are non-deterministic — they don’t always behave the same way, so systematic testing is critical in order to know the system is ready for initial deployment.

For SuplaiCerdas, I wrote three test cases:

  1. Test 1: Price check tool retrieves data from all 3 suppliers
  2. Test 2: Price change calculation detects 17.24% spike as HIGH risk
  3. Test 3: Complete workflow executes end-to-end without errors

Building SuplaiCerdas: The Three-Agent System

Suplaicerdas is comprised of three main agents, as the following:

Agent 1: PriceCheckAgent

  1. Role: Gather Data
  2. Tool: check_component_prices(component_code)
  3. What it does: Queries a simulated supplier database for current prices across 3 suppliers
  4. Sample output:
  5. Supplier_A_Electronics: $5.80
  6. Supplier_B_Components: $6.00
  7. Supplier_C_Global: $5.70

*Actual python code

Agent 2: ComparisonAgent

  1. Role: Risk Analyst
  2. Tool: calculate_price_change(current_price, last_month_price)
  3. What it does: Calculates percentage change and assigns risk levels
  4. High if price increase >10%
  5. Medium if 5–10% price increase
  6. Low if <5% increase
  7. Sample output:
  8. Supplier A: 5.45% increase → MEDIUM risk
  9. Supplier B: 3.45% increase → LOW risk
  10. Supplier C: 0.88% increase → LOW risk

*Actual python code

Agent 3: RecommendationAgent

  1. Role: Procurement Advisor
  2. Tool: Pure LLM reasoning (no external tools)
  3. What it does: Synthesizes price + risk data into a clear, actionable recommendation
  4. Sample output: ‘ RECOMMENDATION: Supplier_C_Global CURRENT PRICE: $5.70 (lowest) RISK LEVEL: LOW (0.88% increase) RATIONALE: Supplier_C_Global offers the best combination of competitive pricing and price stability. Supplier_A_Electronics shows a moderate price increase (5.45%), flagging a MEDIUM risk that could impact your tender margins. ‘

*Actual python code

Testing the solution prototype

Below are the running procedures as well as screenshots of the product:

  1. Run agent test python file, proceed after all test passed.

  1. Run main.py python file and enter component code to be tracked.

  1. After entering desired component code, system will do an automated assessment and return a result of recommendations.

What’s next

As I have achieved a working prototype with a simulated supplier data, there are several directions I can take to further improve this app’s functionality, including the following:

  1. Real Data Integration: Use actual supplier APIs
  2. Currency Conversion: Support different currencies (USD/EUR/IDR)
  3. Automation: Schedule daily price checks with alerts to user for certain conditions like high risk components

Conclusion

Participating in this AI agents course and working on the project has taught me more than just theories. It has taught me how I could implement AI agents into real-life cases, which I believe is crucial during the recent AI developments.

I am excited for what’s ahead.

Thanks for taking the time to read this article :), feel free to leave a comment.

Github project repo link: https://github.com/leonhardsuharjo/SuplaiCerdas_GoogleKaggle-Capstone_AI-Agents_LeonhardSatriaSuharjo

-- written by Leonhard Satria Suharjo --

Top comments (0)