This document summarizes an artificial intelligence project at a liquid cooling manufacturer called EK Water Blocks (EKWB). The goals of the project were to use AI for sales prediction, sentiment analysis, production planning optimization, and a recommendation system. Data scientists built models for sentiment analysis of online forums, sales prediction using ARIMAX methods, and an initial recommendation system using market basket analysis. Future phases aim to prevent out of stock issues and integrate AI further into business processes.
1. Use of Artificial Intelligence in a
Manufacturing Company
Uroš Valant
Using Machine Learning in EKWB company
2. 1
Data Scientists
• CREA pro - Founded in 2010 as a consulting company
• Specializes in Performance Improvement, Business Process Optimization & Data
Science
• Already 15+ successful projects in the field of data science
• Combining business & technology expertise
• Experts in the productive application of R
• Using different data science platforms
2016 NLB
2017
NLB
EKWB
TEKSTINA
3DVA
SBERBANK
ADDIKO
T2
BUDAPEST BANK
RAIFFEISEN BIH
NLB SRBIJA
SBERBANK SRBIJA
2018
3. 2
The Client
• EK Water Blocks – EKWB
• The premium CPU and GPU liquid cooling manufacturer
• The most renowned supplier of high-end premium quality liquid cooling products
• 67% growth in the last two years – 20+ million Euro business
Why AI?
• Vision of their management – the use of future technologies
• Control the growth
• The goal – AI as integral part of daily business
4. 3
€
The Project
Sentiment analysisSales prediction
Recommendation system AI assisted management reporting
First phase
• Upgrade of first four models
• Production planning optimization
• Production stock optimization
Second phase
• Out of stock prevention in sales
• AI in HRM department
• Dynamic pricing
5. 4
Sentiment Analysis – The Scraper
• Special application which harvests text and messages from tech forums and websites
• Runs on the daily basis – scraper as a service
• Extracts texts, clears them and stores them in the database
• Text pre-production, meta data
• Tokenization
I have a MLC 360 on my Predator 360 prefilled 1080 Ti GPU block and my
CPU on my Predator 360 on separate loops.
I'm disappointed the pump doesn't have a standalone PWM header to control
the pump alone. It seems to be integrated with the fans.
6. 5
Sentiment Analysis – The Core
• We conduct the sentiment analysis two times
– Every post from the forum
– For each word from the post
• Syuzhet Package and Stanford dictionary
+ - + -
WEB CONTENT
TOKENIZATON
RECORD SENTIMENT
ANALYSIS
WORD EXTRACTION
WORD SENTIMENT ANALYSIS
WORD CLOUD
7. 6
Sentiment Analysis – The Usage
• Web Application
• Word cloud presentation
• Features:
– Filters
– Search
– Links to original posts
– Alerts
– Sentiment adjustment
– Words statistics
8. 7
Sales prediction
• Forecast integrated in BI reporting
• ARIMAX method
• Comparison with the real time
situation
• Granularity – single product
• Extra events included ( GPU
launches, Black Friday, X-mas)
• Trend prediction
9. 8
Interactive dashboard
• Sales predictions for expert users
• Sales, procurement and production
• Product group lifetime
• Long-term forecast
• B2B and B2C
10. 9
Trickle down effect
• Sales forecast – the top predictive model
• Influence:
– Production
• Planning optimizaton
– Stocks
• Minimal stock requrements
• Out of stock prevention
– Procurement
• Production stock optimization
SALES FORECAST
PROCUREMENT
STOCK
OPTIMIZATION
MINIMAL
STOCK
PRODUCTION
PLANNING
11. 10
Where are we now?
• Recommendation engine
– 60% of sales is conducted through the webshop
– Market basket analysis in the first stage (batch)
– Collaborative filtering in the future (near real time)