Vitalii Bondarenko and Eugene Berko "Cloud AI Platform as an accelerator of enterprise digital transformation"
1. 2018
Cloud AI Platform
as an accelerator of enterprise digital transformation
Vitaliy Bondarenko
Vitaliy.Bondarenko@eleks.com
Eugene Berko
Yevhen.Berko@eleks.com
2. AGENDA
1. Azure AI / ML services
2. Data Processing Architecture Approaches
3. Data Science Platform
4. Lessons learned
3. Office in the USA
New York Office in the UK
London, UK
Office in Eastern Europe
Rzeszow, Poland
Offices in Ukraine
Headquarters
Lviv
Delivery centres
Lviv
Kyiv
Ivano-Frankivsk
Ternopil
a Top 100 Global
Outsourcing
Company
largest IT
companies
in Ukraine
TOP 10
years experience
of delivering
solutions
27
professionals
1200+
ELEKS FACT SHEET
Among the
4. SPEAKERS
Vitaliy Bondarenko
Head of Enterprise Cloud Solutions Office
20+ years of experience; conference
speaker; ELEKS competency manager
and community lead; Solution Architect
for Big Data, Fast Data and AI projects
Eugene Berko
Data Architect
7+ years of experience in BI / Big
Data / DB / DWH. For the last
couple of years has been
developing data solutions of various
nature focusing mostly on high-load
systems and performing enterprise
integration
6. Offering Overview
Tools and technologies:
● Data Science Virtual Machines (both Windows and Linux
based)
● Azure Machine Learning Studio
● Azure Machine Learning Service (preview)
● Azure Batch AI (preview)
Deployment options:
● Azure Machine Learning web service (only for models built
using Azure Machine Learning Studio )
● Python web service in a Docker container
● Apache Spark in Azure HDInsight
● Machine Learning Server (previously Microsoft R Server)
● As T-SQL functions in Microsoft SQL Server
2
7. Azure Cognitive Services
Vision APIs
• Computer Vision
• Custom Vision
Service (Preview)
• Content Moderator
• Face API
• Emotion API
(Preview)
• Video Indexer
Speech APIs
Language
APIs
Search APIs
Knowledge
APIs
• Speech Service
(Preview)
• Custom Speech
Service (Preview)
• Bing Speech API
• Translator Speech
• Speaker
Recognition API
(Preview)
• Bing Spell Check
• Language
Understanding
LUIS
• Linguistic Analysis
(Preview)
• Text Analytics
• Translator Text
• Web Language
Model (Preview)
• Bing News Search
• Bing Video Search
• Bing Web Search
• Bing Autosuggest
• Bing Custom
Search
• Bing Entity Search
• Bing Image
Search
• Bing Visual
Search
• Custom Decision
Service (Preview)
• QnA Maker
9. Stream Analytics Using Native Azure Services
Key points
● Real-time image processing
● Face / objects recognition
● Real-time dashboards for
alerting
● Dashboards for retrospective
analysis and stats
Components
● Computer Vision and Face API
● Event Hub for image injection
● Stream Analytics for
communication with Cognitive
Services
● Blob Storage for storing images
● Cosmos DB for storing model
output
● Power BI for dashboarding /
reporting
10. Batch Analytics Using Native Azure Services
Key points
● Future sales prediction
based on years of data
● Diverse visualization
options
Components
● Data Lake Storage as
scalable storage
● Azure Functions to
transform data into format
more suitable for
machine learning
● SQL Data Warehouse
● Analysis Services to
provide single semantic
model and in-memory
cashing
● Azure SQL Database
● Data Factory
11. Lambda Architecture with Azure Databricks
Key points
● Both streaming and
batch analytics of
online orders
● Anomaly detection
Components
● HDInsight Kafka for
stream injection and
real-time processing
● Azure Databricks as
Apache Spark–based
analytics service with
machine learning
capabilities
● Data Factory for
extracting data and
injection into main
storage
13. AI Solutions
Challenges:
• How to feed data to the AI model?
• How to control access to output of the
models that can contain critical business
information?
• Hot to react to increased data velocity?
• How to deploy models to production
environment?
• How to retrain models in an efficient way
without data scientists?
• How to measure actual effectiveness of
the model on actual data?
• How to integrate with existing enterprise
infrastructure?
• How to make instant decisions according
to AI predictions?
14. Data Science Platform
Target Customers
Enterprises in the state of digital transformation
which are building strategies of AI and Big Data
implementations.
Key points in solution vision:
• Real-time analytics and lightning-fast response to
incoming data no matter how big it is
• Removing the pain of model management and
deployment from developers
• Easy model scaling for both scoring and training
15. Trained Models Registry and Deployment
Registry of trained models
● Metadata for Models
● Versions
● Unified UI for Deployment
Deployment
● Create pod on Kubernetes
● Flask for Python
● POJO unified model
● Schema Registry
Monitoring
● Scoring Statistics
● Automatic Validation
16. Visualisation for Anomaly Detection and Real-Time Scoring
Capabilities
● Machine Learning models training on historical data
● Real-time models scoring
● Integration with Enterprise applications
● Real-time Data Visualisation
Real-time Machine Learning
● Shopping Behaviour Analysis
● Logs Anomaly Detection
● Fraud Prediction
● Product Recommendation
● Campaign Recommendation
● Demand Prediction
● Route Optimization
● Customer Segmentation
Benefits
● Expert controlled model training
● Validation jobs for all models
● UI for models deployment and monitoring
● Latency in 1 second
17. Deployment and Scalability with Docker and Kubernetes
Capabilities
● Deployment to Cloud and On-Premises
● Docker containerization
● Kubernetes Cluster
● Automated continuous integration
● Unified cluster for all Platforms
● UI for Cluster Management
Benefits of Kubernetes
● Scalable on level of VMs
● Integration with Enterprise Network
● Enterprise Level Security
● REST API
● Platform for Model deployments
20. Lessons Learned
Key points
● Azure is a mature environment for Data
Engineering, Machine Learning, and Platform
Building
● HDInsight is a powerful Hadoop-based System
for real-time and batch data processing
● Cosmos DB is quite sophisticated data base
and needs more panels for configurations
● AKS is an excellent tool for microservices and
better than native Kubernetes
● PowerBI is very helpful for real-time analytics
● Databricks is a power Data Platform and has
bright future.