DXC Industrialized A.I. – Von der Data Story zum industrialisierten A.I. Service

1. © 2017 DXC Technology Company. All rights reserved.
DXC
Industrialized A.I.
May, 2019
From data story to industrialized
A.I. service

2. May 10, 2019 2
• AI, How it Starts
• Map the AI Journey
• DXC’s Industrialized AI
Agenda

3. © 2017 DXC Technology Company. All rights reserved.
How it Starts
The journey begins

4. May 10, 2019 4DXC Proprietary and Confidential
PD_9991a-19
Cognitive Computing
Simulating, specifically, the perception and reasoning
aspects of human intelligence:
• Natural-Language Processing
• Speech
• Vision
Artificial Intelligence
Any program that does
something that we would think of
as intelligent in humans
AI is defined by the application of
the technology rather than the
technology itself. What is considered
AI may change over time.
AI at DXC:
• Extend Domain Expertise
• Perform Complex Planning
• Infer intent
Unsupervised Supervised
Discover new patterns Learn specific patterns
Cat
Not Cat
Artificial Intelligence
Machine Learning
Any program that improves
its performance through
experience rather than
explicit programming.
Deep Learning
Machine learning based on
neural networks

5. May 10, 2019 5DXC Proprietary and Confidential
PD_9991a-19
Strong versus weak A.I.

6. May 10, 2019 6DXC Proprietary and Confidential
PD_9991a-19
AI or ML?
Face detection Infer one is upset
Schedule a repair
(Predictive Maintenance)
Predict equipment breakdown
Sort documents by
topic (e.g. emails)
Cluster documents by similarity
AI
ML AI
ML
AI ML

7. © 2017 DXC Technology Company. All rights reserved.
How get to
Enterprise Scale
A.I. Experience?
The first steps

8. May 10, 2019 8DXC Proprietary and Confidential
Enterprise-Scale
Data Science
Experience
Industrialized AI
Master
The Industrialized AI
Master journey

9. May 10, 2019 9DXC Proprietary and Confidential
Create Data Stories
Run Agile
Transformation
Industry Consulting Experience
Industrialized AI
Leader
The Industry Consultant

10. May 10, 2019 10DXC Proprietary and Confidential
Enterprise-Scale
Data Science
Experience
Industrialized AI
Master
Create Data
Stories
Run Agile
Transformation
Industry Consulting
Experience
Industrialized
AI Leader
The Industry Consultant
Journey

11. May 10, 2019 11DXC Proprietary and Confidential
A Common Mistake with AI Projects
• Without a hypothesis: convincing stakeholders to take the leap
Data Write
algorithms
Find
patterns
Tell a
story
Stakeholder
commitment
Big Gap

12. May 10, 2019 12DXC Proprietary and Confidential
From Ideas to Innovation
Buildathons are a great way to begin innovating with AI. We have created some effective formats to make
sure the best ideas are transformed into finalized products ready to be launched.
1
2 weeks
Preparation
2
48 hours
Buildathon
3
3 months
Industrialize
For 2 weeks leading up to a buildathon,
the preparation phase allows us to
collect data, ideate and refine ideas.
A post-buildathon phase where we
industrialize the solution through a
carefully structured AI innovation
program.
Teams composed of data scientists,
data engineers and analytics developers
create an AI solution in 48 hours around
a theme.

13. May 10, 2019 13DXC Proprietary and Confidential
Various skill sets of the DXC team
AI Leaders
Those who have sizzling ideas and
are looking for a team. They have
a vision of how AI can transform
the company.
Their talents include building,
testing and analyzing AI solutions.
Love collaboration and innovation.
Experts in munging data and
building analytics platforms. They
know how to scale AI and make an
impact.
They know how to build data-
driven apps that reach employees,
and change how business is done.
The secret ingredient to any team.
Data Scientists Data Engineers Analytics Developers

14. May 10, 2019 14DXC Proprietary and Confidential
PD_9991a-19
Analytical Layer Information Layer Operational Layer Benefit Layer
Ticket classification
Topic-based index system
Incident Ticket
Handbooks
NLP pipeline
Automatic ontology/
topic assignment
Optimized ServiceInstruction article +
recommendations
Solutioning support
Savings
Quality + Speed +
Efficiency Increase
Feedback loop
Customer
Satisfaction
50M IT tickets issued
At the forefront of technology, ITSM is
heavily impacted by the rapid change of
the IT landscape
10% less time spent
Assigning service tickets according to
their identified topic to specific teams,
obsoletes manual distribution.
10k pages of manuals read
Handbooks for knowledge manag
ment continue to be the
primary resource to go to
NLP
>95%classification accuracy
State of the art Deep Learning
classification algorithms achieve
highest scores in real-time
Digital Assistance for services like maintenance / other services
Enhance/ improve
ontology/ corpus
20% more efficiency
Ontologically indexed Knowledge Base
articles and article recommendations
boosts ticket-team productivity
15% cost savings
By cutting out inefficient workflow steps
through automation and efficiency gains
through digital assistance
25% more satisfaction
Faster ticket resolution and the
increased quality leads to increased
customer satisfaction
Version & Release Date index
System.
Feedback loop

15. May 10, 2019 15DXC Proprietary and Confidential
PD_9991a-19
What Clients See (The AI Market)
Platform
Product
“Solution”
Platform
Product
Platform
Product

16. May 10, 2019 16DXC Proprietary and Confidential
Enterprise-Scale
Data Science
Experience
Industrialized AI
Master
Run AI
experiment
Perform AI
Forensics
Machine Learning
Experience
Industrialized AI
Data Scientist
The data science
journey

17. May 10, 2019 17DXC Proprietary and Confidential
Approach using Python and RASA.io
Dataset
Questions, Answers, Links,
Tags, Importance, …
e.g. mechanics.
stackexchange.com
Dataset Data Preprocessing
From Posts.xml to a
cleaned and preprocessed dataframe
User
Feedback Loop
Machine Learning &
Application Development
Develop a Chatbot with Intent Recognition
with the aid of Natural Language
Processing & Understanding

18. DXC Proprietary and Confidential May 10, 2019
Industrializing AI

19. May 10, 2019 19DXC Proprietary and Confidential
Seed Stage
• Curate the idea
Early Stage
• Solve the problem
Growth Stage
• Perfect the data supply chain
Maturity
• Automate the infrastructure
The AI Garage
• Virtual meetups
• Virtual Build-a-thons
Industrialized
Co-located
Build-a-thons
Focused Sprints Managed Services
Industrialized Industrialized
Iterations 0.X
functional
operational
managed
functional
operational
managed
functional
operational
managed
Iterations 1.1…1.X
Iterations 2…N
Give the Client an AI Startup Experience
Crawl Walk Run
Investment:
Participation
Investment:
~$100K - $400K
Investment:
~$0.5M-$1M
Investment:
~$5M

20. May 10, 2019 20DXC Proprietary and Confidential
Industrialized AI Strategy: Use AI to Open New Innovation Capacity
Genesis Custom built Product Commodity
Visibility

21. May 10, 2019 21DXC Proprietary and Confidential
Industrialized AI Strategy: Use AI to Open New Innovation Capacity
Genesis Custom built Product Commodity
Visibility
Mature, stable commodities

22. May 10, 2019 22DXC Proprietary and Confidential
Industrialized AI Strategy: Use AI to Open New Innovation Capacity
Genesis Custom built Product Commodity
Visibility
Highly
visible in
the
enterprise

23. May 10, 2019 23DXC Proprietary and Confidential
Enterprise-Scale
Data Science
Experience
Industrialized AI
Master
Build Utility AI
Services
Build data
pipelines
Industrialized AI
Data Engineer
Data Engineering
Experience
The data engineering
journey

24. May 10, 2019 24DXC Proprietary and Confidential
Data Science is OSEMN
You are awesome.
I am awesome.
Data Science is OSEMN.
OSEMN Pipeline
•O —Obtaining our data
•S — Scrubbing / Cleaning our data
•E —Exploring / Visualizing our data will allow us to
find patterns and trends
•M —Modeling our data will give us our predictive
power as a wizard
•N —Interpreting our data

25. May 10, 2019 25DXC Proprietary and Confidential
Obtain Your Data
Skills Required:
•Database Management
•Querying Relational Databases
•Retrieving Unstructured Data
•Distributed Storage
Extract the data into usable format

26. May 10, 2019 26DXC Proprietary and Confidential
Scrubbing / Cleaning Your Data
Objective:
•Examine the data: understand every feature you’re working with, identify errors,
missing values, and corrupt records
•Clean the data: throw away, replace, and/or fill missing values/errors
Skills Required:
•Scripting language
•Data Wrangling Tools
•Distributed Processing

27. May 10, 2019 27DXC Proprietary and Confidential
PD_9991a-19
Build and Manage Industrialized Data Pipelines
REST
API
Event
Queue
Establish
automated,
continuous and
secure access
to source data
REST
API
Streaming
data
REST
API
Realtime
Analytics
Batch data
REST
API
File/Data
store
REST
API
Batch
Analytics
Data AnalyticsData Engineering
Maintain a
comprehensive
set of data
pipelines
needed to
create and
validate
actionable
insight
Integrate data and
use automation to
maintain global
context
Match data to
expected format,
structure, schema,
and content
We work with
clients to…
Enrich data with
additional features
(and AI insights) that
increase the ability to
predict target
outcomes
REST
API

28. May 10, 2019 28DXC Proprietary and Confidential
PD_9991a-19
Real-Time Data
Historical
Ticket Dataset
Streaming
Ticket Data
AWS
CodeBuild
RASA
Runtime in
AWS ECS
Back-End
Inference
Parse text,
Determine intent
Real-Time
Data
Trained NLP Model
Performance MetricsBatch Data
Real-time ticket
intent prediction
Web Application
for analysis, visualization,
reporting of NLP data
NLP
predictions
AWS S3
Storage
Training Set
Test Set Train
Generate Model,
measure effectiveness
AWS
CloudWatch
Target
System
Current Model
AI Utility – Actual Deployment
Hypothesis:
We can use NLP technology to process service requests and determine the maintenance intent
Ingest Data Pipeline
Deploy and Secure
Analyze, Design, Train, Test, Score

29. May 10, 2019 29DXC Proprietary and Confidential
Exploring (Exploratory Data Analysis)
Objective:
•Find patterns in your data through visualizations and charts
•Extract features by using statistics to identify and test significant variables
Skills Required:
•Python
•R
•Inferential statistics
•Experimental Design
•Data Visualization

30. May 10, 2019 30DXC Proprietary and Confidential
Modeling (Machine Learning)
Objective:
•In-depth Analytics: create predictive models/algorithms
•Evaluate and refine the model
Skills Required:
•Machine Learning: Supervised/Unsupervised
algorithms
•Evaluation methods
•Machine Learning Libraries: Python (Sci-kit
Learn) / R (CARET)
•Linear algebra & Multivariate Calculus

31. May 10, 2019 31DXC Proprietary and Confidential
Best Practice: An Incremental, Agile Approach
Stakeholder
commitment
Hypothesis
Get data
Write algorithmsGenerate
evidence
Decide on
hypothesis credibility
✓
Take an action
Data science
Small 4-6 week sprints
Scale globally across
the enterprise
and adapt
to fluctuating
enterprise demand
Map to standard
concepts and
make insights
repeatableUse experiments to
produce reliable
measurable results
Produce insights
that can be
distributed and
used throughout
the enterprise
Data engineering

32. May 10, 2019 32DXC Proprietary and Confidential
Interpreting (Data Storytelling)
Objective:
•Identify business insights: return back to business problem
•Visualize your findings accordingly: keep it simple and priority driven
•Tell a clear and actionable story: effectively communicate to non-technical audience
Skills Required:
•Business Domain Knowledge
•Data Visualization Tools
•Communication: Presenting/Speaking &
Reporting/Writing

33. May 10, 2019 33DXC Proprietary and Confidential
AI - 4. Industrialize your AI
A. Operationalization (Managed Platform and Managed Security)
Use Case: productionize Sandbox-environment
B. Industrialization (AI Utility, Closed AI Loop)
Use Case: End-to-end Automation and Scalability
-> Example „Knowledge Management (KM) Article Prediction Mechanism”
a. Reduce human effort
b. Reduce incident resolution time
c. Enhance knowledge management
d. Enhance consistency of incident resolution
Results:

34. DXC Proprietary and Confidential May 10, 2019
AI Utility Process Flow & Model Management
Labeling Team
Created Arthur Shlain
romthe Noun Pro ect
NLP for
Intent Recognition
6) Determine
intent Test Results
1) Parse text with NLP
2) Create
Categorization Model
Tags.
Synonyms
3) Train Model
4) Validate
Categorization
Model
5) Operationalize
Model
7) Route to L2
support
Historical DB
with questions
and intent
Data Scientists

35. May 10, 2019 35DXC Proprietary and Confidential
PD_9991a-19
Real-Time Data
Historical
Ticket Dataset
Streaming
Ticket Data
AWS
CodeBuild
AWS ECS
(Docker)
Back-End
Inference
Parse text,
Determine intent
Real-Time
Data
Trained NLP Model
Performance MetricsBatch Data
Real-time ticket
intent prediction
Web Application
for analysis, visualization,
reporting of NLP data
NLP
predictions
AWS S3
Storage
Training Set
Test Set Train
Generate Model,
measure effectiveness
AWS
CloudWatch
Target
System
Current Model
A Simple NLP Pipeline using Rasa NLU
Hypothesis:
We can use NLP technology to process service requests and determine the maintenance intent
Only this part used for the current badge
pipeline:
- name: "intent_featurizer_count_vectors"
- name: "intent_classifier_tensorflow_embedding"
intent_tokenization_flag: true
intent_split_symbol: "+"

36. May 10, 2019 36DXC Proprietary and Confidential
PD_9991a-19
• Make entry into the
cloud–with either
Amazon AWS,
Microsoft Azure, IBM
or HPE Helion Virtual
Private Cloud (VPC)–
uncomplicated and
unintimidating,
minimizing costly
learning steps through
a high-touch service
approach, not widely
available
Why choose DXC Technology?
• For production in a
hybrid model, DXC
advises and implements
deployments with a
range of options,
including HPE Helion
VPC, on-premises and
public cloud
environments.
• Our breadth of expertise
and methods provide
options few competitors
offer
• Get accelerators like
reference architectures
and deployment
automation, extended
analytic capability
options, blueprints and
runbooks that cover the
initial setup,
onboarding and
ongoing run with SLAs
• We provide the richest
standardized package
in the industry
• Best practices for analytic
applications and data
workload optimization –
DXC combines the long
term commitment to
business intelligence (BI)
and data management with
access to a broad variety
of real life cases.
• We have proven expertise
managing Hadoop, related
analytic technologies and
cloud native services for
enterprise solutions
Full service
enablement
Best Practices
and Expertise
Easy and safe
setup
Integration and
expansion
options

37. May 10, 2019 37DXC Proprietary and Confidential
Enterprise-Scale
Data Science
Experience
Industrialized AI
Master
Create Data
Stories
Run Agile
Transformation
Industry Consulting
Experience
Industrialized
AI Leader
Build Utility AI
Services
Build data
pipelines
Industrialized AI
Data Engineer
Data Engineering
Experience
Run AI
experiment
Perform AI
Forensics
Machine Learning
Experience
Industrialized AI
Data Scientist
DXC Industrialized AI
The journey has begun
Learning happens everywhere

38. May 10, 2019 38
Where are you on the journey to
industrialized AI?
Contact us for a free copy of the booklet

39. DXC Proprietary and Confidential
Thank you.

40. DXC Proprietary and Confidential
Sources:
• https://blog.openai.com/better-language-models/
• https://d4mucfpksywv.cloudfront.net/better-language-
models/language_models_are_unsupervised_multitask_learners.pdf
• http://www.informatik.uni-oldenburg.de/~iug08/ki/Grundlagen_Starke_KI_vs._Schwache_KI.html
• https://towardsdatascience.com/overfitting-vs-underfitting-a-conceptual-explanation-d94ee20ca7f9
• https://deepmind.com/blog/alphafold/
• https://towardsdatascience.com/a-beginners-guide-to-the-data-science-pipeline-a4904b2d8ad3
• https://openai.com/blog/ai-and-compute/