[DSC DACH 24] How to optimize stock with an Automated Logistics and Intelligent Supply chain Assistant - Fabio Eupen

Optimizing supply chain and stock management is crucial for increasing sales while at the same time wasting less resources and reducing costs. In this talk we will go through the key challenges that need to be addressed in supply chain optimization, in order to increase the availability of the right product in the right place, at the right time. We will outline the essential building blocks for successfully implementing an automated logistics and intelligent supply chain assistant using machine learning algorithms, demonstrating how it supports planners in decision-making. Finally, we will explore how generative AI can enhance demand forecasting, improve order recommendations, and provide explanations of the results to the users. Attendees will gain insights into innovative approaches and practical solutions for modern supply chain management.

1. How to optimize stock with an Automated Logistics and Intelligent Supply chain Assistant Fabio Eupen 13.09.2024 DATA SCIENCE CONFERENCE DACH 24

2. How to optimize the supply chain? How to improve with generative AI? How to implement successfully?

3. How to optimize the Supply chain?

4. Scheduling purchase orders according to forecasted demand Schedule work orders to satisfy market demand Supply Chain The challenges

5. Scheduling purchase orders according to forecasted demand Schedule work orders to satisfy market demand Send items to retail stores and wholesale customers Supply Chain The challenges

6. Scheduling purchase orders according to forecasted demand Schedule work orders to satisfy market demand Send items to retail stores and wholesale customers Replenishment between stores Supply Chain The challenges Have the right product at the right place at the right time Avoid stock-outs & overstock

7. Supply Chain The challenges Have the right product at the right place at the right time Avoid stock-outs & overstock 1| DEMAND FORECASTING 2| AUTOMATED & OPTIMISED ORDERS 3| CARGO & LOGISTICS OPTIMISATON 4| PRODUCTION PLANNING OPTIMISATION 5| STORE TRANSFERS

8. The building blocks for optimization Demand forecasting algorithm based on historical data, 3rd pa r ty data and all relevant constraints for each country / channel / SKU. Includes options of promotion planning and pricing . | Demand forecasting Modul dedicated for a replenishment between stores in complex retail environment. To send items to stores with maximum chance of sale , while keeping proper assortments. System directly integrates with existing ERP system via APIs, thus is not another system to maintain.. Order optimization algorithm providing user optimal quantity to order based on all relevant constraints. Works for multiple locations/channels and includes cargo optimization . | Order recommendation Interactive reports present centralised inventory management platform within Qlik or Power BI . Evaluation includes standardised KPI’s and metrics to track system performance. | KPIs and reporting Planning o ptimization algorithm will schedule work orders in optimal order to maximize output in shortest time period, while satisfying all relevant constraints . | Production/manufacturing planning | Store transfers/replenishment Alerting module for real - time alerting in case of out-of-stocks or near out-of-stock , and overstock , to timely plan production and promotions to maximise efficiency . | Alerting GPT-powered assistant for any user question , regarding the platform, forecasts, recommendations or further information about the products . Improve forecasts through GPT-generated features | Virtual assistant A utomated L ogistics and I ntelligent S upply chain A ssistant | Integration APIs 1 | 2| 3| 4| 5|

9. How to improve with GenAI ?

10. Recap of the Replenishment automation solution DEMAND FORECASTING INVENTORY MONITORING RECOMMENDING PURCHASE ORDERS ORDER APPROVAL AND SCHEDULING

11. DEMAND FORECASTING INVENTORY MONITORING RECOMMENDING PURCHASE ORDERS ORDER APPROVAL AND SCHEDULING Enhanced with GPT DATA SOURCES Supplier communication Shipment tracking APIs Supplier databases Logistics data News articles Social media INSIGHTS Supplier issues Product shortages Shipment delays Viable Alternatives Automatically (scheduled) Notification & chatbot NOTIFYING STAKEHOLDERS GPT based chatbot lists and explains the recommendations as well as any other part of the process to the stakeholders. 

12. Grounding via Retrieval Augmented Generation

13. How chatbot explains data and results Stock related Convert question to SQL statement via GPT Get data from database via SQL statement Add relevant data to prompt Forecast related Access parameters of forecast Access features and feature importances of forecast Recommendation related Access the underlying rules Access additional insights GPT answers original question with data enriched prompt GPT decides which data is needed User question “Which product sold best in the last three months?” “Why is product B forecasted to sell much less in the next week?” “Why don’t we order more of product C?” Chatbot reply “Product A sold 42,000 units in the last 3 months.” “Product B’s sales are highly weather dependent and next week will be cold and rainy.” “Because product C is not available for the next two months, but we are ordering product Z as an alternative.”

14. Adding GPT generated features to the forecast Machine Learning approach: Sales data + Features  Sa les forecast t t today 100 200 300 400 ML m odel True Value Average Statistical methods Sales in last period Last year Sales Promotional activities Calendar & Events Weathe r DATA Economic situation News features Customer feedback GPT enhanced features:

15. How to implement successfully?

16. Key factors for a successful implementation Data suitable for ML project and of good quality Leadership and main stakeholders are convinced of importance of improved digitalization In-depth requirement analysis with well documented objectives, limits, and timeline Involve the main users from the very beginning including the test and validation phase Integrate with existing (ERP) systems Visualize KPIs and metrics in a tool of the customers choice Be aware of industry specific customizations Constant feedback from key users and stakeholders

17. Conclusion (after a successful implementation) Happier workers (easier decision making through forecasts and suggestions) Happier customers (desired product more often directly available in the right place) Increased profit (decreased number of stock outs) Increased level of flexibility (reduced production time) Less needed storage capacity (decrease of stock level) Reduction of waste (less production of non-selling items) Increased cash flow (less cash bound in materials and products) Less money spent on new machines (better utilization of existing machines) Reduction of traffic and emissions (less and optimized shipment)

18. Thank you! Fabio Eupen [phone] [email protected]

19. Thank you! Fabio Eupen [phone] [email protected]

20. Backup Slides

21. DIFFERENT DATA SOURCES CLEAN DATA BUILDING FEATURES DIFFERENT ML MODELS ERP, BI, CRM, external data True and relevant data Outliers , stock-outs , promotions Substitute data Date & sales fe a tures Contextual features External features Different algorithms Granularity - daily , weekly , monthly Forecasting per item & location How does it and why does it work

22. How does it and why does it work DIFFERENT DATA SOURCES CLEAN DATA BUILDING FEATURES DIFFERENT ML MODELS ERP, BI, CRM, external data True and relevant data Outliers , stock-outs , promotions Substitute data Date & sales fetures Contextual features External features Different algorithms Granularity - daily , weekly , monthly Forecasting per item & location

23. Forecasting ( r egression) Machine Learning approach: Sales data + Features  Forecast sales t t today 100 200 300 400 ML m odel True Value Average Statistical methods Sales in last period Last year Sales Promotional activities Calendar & Events Competition Weather ,… DATA

24. Production Planning

25. Evolutionary Algorithm Solution 1 Solution 2 … Parameter 1 Parameter 2 … Population Selection Crossover Mutation Evaluation Objective function ranking 2 3 1 4 Selection Replacement

26. Cargo Optimization Minimize number of trucks by best fitting available items by their volume and weight constraints If trucks full, prioritize which items go first If trucks still have space, predict for how long waiting can be beneficial [explain better!] A A A B A C C B A A A C  A C C C C C A A A B A C C B C A A A C C C C

27. Discriminative AI versus Generative AI with

28. Grounding Techniques to make LLMs more Reliable Grounding is the process of adding information to the pre-trained LLM that might be specific to the use case. The simplest form is by manually adding context to the prompt. Retrieval Augmented Generation (RAG) Model stays the same Prompt (system message) gets enriched by relevant external information (internet, internal database, pdfs, documents,…) Forces model to only answer with this information which strongly reduces the probability of hallucinations Fine-tunning Re-training the model and changing the underlying neural network Input: example questions and answers Less flexible than RAG, e.g., the whole model needs to be re-trained if just one document is added More prone to hallucinations

29. Enhance the Replenishment automation product in 3 ways : Improvement of system’s recommendations through automation of gathering and analysis of external data S upply-chain disruptions , vendor unavailability and stability Enhancement of how users interface with the system A sk for explanations Lowering the technical barriers between the human inventory planner and the complex AI system E stablishing natural language as the method of communication Purpose & Benefits

30. Supply chain The challenges 1| DEMAND- FORECAST 3| PURCHASE ORDER SCHEDULE PRODUCTION PLANNING OPTIMIZATION 2 | * CARGO- OPTIMIZATION *

31. How chatbot explains data and results Stock related Convert question to SQL statement via GPT Access data from database via SQL statement Forecast related Access parameters of forecast Access features and feature importances of forecast Recommendation related Access the underlying rules Access additional insights GPT answers original question with data enriched prompt GPT decides which data is needed User question “Which product sold best in the last three months?” “Why is product B forecasted to sell much less in the next week?” “Why don’t we order more of product C?” GPT reply “Product A sold 42,000 units in the last 3 months.” “Product B’s sales are highly weather dependent and next week will be cold and rainy.” “Because product C is not available for the next two months, but we are ordering product Z as an alternative.” UI related Create embedding of summary of the question and chat history Get relevant data from documentation via embedding vector data base

32. The building blocks for optimization Demand forecasting algorithm based on historical data, 3rd pa r ty data and all relevant constraints for each country / channel / SKU. Includes options of promotion planning and pricing . | Demand forecasting Modul dedicated for a replenishment between stores in complex retail environment. To send items to stores with maximum chance of sale , while keeping proper assortments. System directly integrates with existing ERP system via APIs, thus is not another system to maintain.. Order optimization algorithm providing user optimal quantity to order based on all relevant constraints. Works for multiple locations/channels and includes cargo optimization . | Order recommendation Interactive reports present centralised inventory management platform within Qlik or Power BI . Evaluation includes standardised KPI’s and metrics to track system performance. | KPIs and reporting Planning o ptimization algorithm will schedule work orders in optimal order to maximize output in shortest time period, while satisfying all relevant constraints . | Production/manufacturing planning | Store transfers/replenishment Alerting module for real - time alerting in case of out-of-stocks or near out-of-stock , and overstock , to timely plan production and promotions to maximise efficiency . | Alerting GPT-powered assistant for any user question , regarding the platform, forecasts, recommendations or further information about the products . Improve forecasts through GPT-generated features | Virtual assistant A utomated L ogistics and I ntelligent S upply chain A ssistant | Integration APIs 1 | 2| 3| 4| 5|

33. Adding GPT generated features to the forecast Machine Learning approach: Sales data + Features  Forecast sales t t today 100 200 300 400 ML m odel True Value Average Statistical methods Sales in last period Last year Sales Promotional activities Calendar & Events Weathe r Economic situation DATA News features Customer feedback GPT enhanced features:

34. Adding GPT generated features to the forecast Machine Learning approach: Sales data + Features  Forecast sales t t today 100 200 300 400 ML m odel True Value Average Statistical methods Sales in last period Last year Sales Promotional activities Calendar & Events Weathe r Economic situation DATA News features Customer feedback GPT enhanced features: