Artificial Intelligence is real and over the next few years, will significantly change the world as we know it. Even though there is some hype around this technology, many companies have put this to practical use, helping businesses delight their customers, increase engagement on their Platforms, and ultimately delivering positive financial results. Many companies have already invested into IT Platforms and it is not practical to invest in bringing up an Artificial Intelligence Platform in parallel. So, the more practical approach is to add Artificial Intelligence capabilities to existing Platforms. In this keynote session, Harish Nalagandla, Director of Engineering, Enterprise Services Platform, PayPal, will discuss the pillars that form the foundation of Artificial Intelligence and will describe an approach towards how companies with existing systems can add Artificial Intelligence capabilities to their existing Platforms. This session will present a high level blueprint to help guide organizations in their own journey towards adoption of Artificial Intelligence technologies as they make progress towards digital transformation. This is a high level presentation targeted towards technology executives.

1. Harnessing the Power of Emerging
Technologies
An approach towards adding AI Capabilities to existing Platforms

2. Agenda
2
AI Overview
What makes AI possible Now?
Deep Dive on
Data | Training | Inferencing
Bringing It All Together
Cloud
People
Q&A

3. AI in the industry and at PayPal.
• Some examples of AI @ PayPal
• PayPal Fraud Prevention
• PayPal OneTouch
• PayPal Customer Service
• Intelligent Routing of Calls
• Determine Best Next Action
• Self Help. Chat Bots.
Automation
Internet and cloud
Health Care
Media
Security and Defense

4. What makes AI possible, now?
4
Big Data Training Inference
High Performance Compute / GPU

5. Spam Detection – Let’s take Spam Detection as a use case..
5

6. Data – Deep Dive
6
Identify and prepare your data for training
Identify Data Needs and
Collection Points
Data Processing and Persistence
Data Access
Data Management
Data
warehous
e
K/V
Datastore
s
Existing
RDBMS
TimeSerie
s Data
stores
System
/Server
Logs
Smart Data
import
Messaging
Platforms (MQ,
Kafka)
Data
Connectors
Data Catalog
Data Quality and
Availability
Metrics
Building Blocks
People
Lifecycle of Data
Platform EngineersData Engineers Infrastructure Engineers
Data Stores
Pipes
Tools

7. Data Lifecycle – Architecture recommendations
7
Source: https://www.paypal-engineering.com/2018/04/17/gimel/

8. 8
Lifecycle of Training
Training UI
Feature Engineering
Model Training
Model Testing
Model Deployment
Jupyter
Notebooks
Data Catalogs
Data
Transformatio
n
Support for
"Model Zoo"
Support for
various
training Tools/
Libraries.
Training jobs
deployable to
GPUs
Output in formats
that ONNX,
Protobuf etc.
Building Blocks
People
Training – Deep Dive
How to take raw data, extract insights and build models that can be used for real time
decisioning
References Model Zoos:
https://github.com/caffe2/models/ | https://mxnet.apache.org/api/python/gluon/model_zoo.html | https://github.com/apache/incubator-mxnet/tree/master/example
Infrastructure EngineersData Scientists

9. Training and Inference – Ecosystem
9
Cognitive Toolkit

10. 10
Lifecycle of
Inference
Building Blocks
People
Inference – Deep Dive
Use the model for real time, predictive decisioning
Platform EngineersData scientists Infrastructure Engineer
Model deployment lifecycle
Model execution
Platform –ilities
Inference Results fed back to
Training.
Model performance evaluation
Model
Deployment
Pipeline
Design as
RESTful /
Microservices
decision
applications
Integration with
your Real Time
Decisioning
Feedback Loop,
decisions made
become events
that are fed back
to the Training
System ‘-ilities’
are an important
consideration.
Decision
Monitoring and
Model Score
monitoring.

11. Inferencing – Deep Dive(RT, NRT and Batch)
11
Real time Near Real
time

12. Bringing it all together
12
Data Training Inference
High Performance Compute / GPUs

13. Cloud Strategy
13
Public Cloud Private Cloud Hybrid
Training
Data
Inference
HPC
Partial workload on the cloud,
part in your own data centers
Public Cloud Providers to manage
your data center (you are the
only tenant of that Cloud)
There are a few options to deploying this whole pipeline. Pick the best option based on your
need.
If starting new, start on the public cloud and move parts to private cloud as needed.

14. People Strategy
14
• Last but not the least, it is important
to pull together the right teams with
the right skills to build these AI
capabilities.
• Other key roles like Product Owners,
Architects, Program Managers,
Engineering Leaders are needed too.
• The roles described here are critical
skills and are not that easily
available in the industry.
• Hire the right people with right skills
as a foundational team.
Infrastructure
Engineers
Data Scientists
Platform Engineers
Data Engineers
It’s all about the people!

15. Q & A