AIC – CS DevScript

Thank you Bhavesh.

Hello All, I am Dev Pathi, Chief Technology Officer at Credit Saison, India

About Credit Saison Global –

Before talking about how we leveraged Amazon Sagemaker service for our infrastructure, let me quickly give you a brief about Credit Saison.

Credit Saison is a Japan based NBFC headquartered in Tokyo. They are one of the largest credit card providers in Japan having established the business since 1951. They have been expanding globally (APAC in particular) over the past 10 years. Today, they have subsidiaries in Singapore (Investments), Indonesia, Philippines, Vietnam (Debt Partnerships with large local banks) amongst others.

Credit Saison India

Credit Saison India is a subsidiary of its Japanese parent We setup the Indian entity 4 years back with the intention of being a Technology-led NBFC. So, while the Japanese parent company is an issue of credit cards, the Indian entity serves the local market providing debt/loans. In India, we are currently a AAA rated NBFC and serve both – Retail and MSME segments by providing debt.

Fast forward to today, we have successfully deep routed ourselves into the eco-system with a substantial loan book size powered by our in-house tech. Our USP has been about integrating and scaling on digital partnerships and has given us the momentum to go direct to market (where we need plenty of AI/ML).

We have over 400 employees across India of which the tech team alone is close to a 100 folks. We have also engaged with CloudThat, a technology partner on a few models – providing some Advanced Technical Support and Monitoring of our systems & importantly in this context, providing solutions to improve our Data/AI/ML stack which are deployed on AWS.

We will be discussing the MLOps and its features that was used in the upcoming slides.

Cloud Journey

Our philosophy has been to be more than just another fintech organization. We are keen to make strong financial impacts and our inception strategy was to ensure the building of robust, scalable and responsible tech.

With no legacy set-up and to take a digital-first approach, we obviously had our eyes set on cloud adoption from the start. This mindset helped pave the way for having conversations on scalability and risk minimization.

Our technology is built to address digital partnerships; however, the stack model was built with the mindset of any financial service domain over a generic platform. Of course, we concentrated some of this domain mindset towards lending and partnerships. After a successful ongoing stint, we had enough courage to launch our own products directly in the market. So, while not exactly remodeling, we did improvise on our overall tech stack. For us to launch directly in market, we need to make decision from our previous expertise by relying on the data we have accumulated over the years and applying our learning to make credit worthy decisions in real time.

Our inception on Cloud with AWS

Post considering the cloud providers options available in Indian market, key factors that prompted the move to AWS. Major drivers for us to choose AWS were — the familiarity of service offering, flexibility in model, developer friendliness, history of reliability and innovation and business support. I have had experience with AWS in my previous organization. Of course, it helped to have the relationship and comfort from the start and the ease of explaining the task we were setting up with Credit Saison in India. We have received a solution-first approach with suggestions on the right service offering, pricing, the choice of Partners like CloudThat.

Security Benefits post AWS

Considering the segment, we belong to, and the involvement of money, Security is an extremely critical aspect of our DNA. Setting up policies from Day 1 and continually improving on them has helped us safe till now.

While examining security, we must consider internal as well as external risk factors. And this applies while architecting the solution as well. We have always stressed smart access controls, whether it is for accessing and setting up cloud tools to authenticate and authorize the APIs or even how the data can reside on an ephemeral or static layer.

There are a bunch of AWS services that have come to our rescue in getting our functionality and IT costs optimized. To name a few

Compute
Amazon EC2s has been our go-to compute service. It provides resizable computing capacity in the cloud. The reduced time taken for scaling up capacity quickly minimizes the cost by a lot.

AWS Lambda has been used across applications for multiple background processing

And now Amazon SageMaker which will take our ML models from development to production

AWS Elastic Load Balancing

Database

Amazon RDS

Amazon DynamoDB

AWS Database Migration Service

Amazon Redshift

Storage

Amazon S3

Amazon Elastic Block Store

Web and Mobile + Serverless

Amazon SNS

AWS API Gateway

Amazon Cognito

Amazon SQS

AWS Step Functions

AI/ML

Amazon Rekognition

Amazon Sagemaker

Governance

AWS Secrets Manager

Amazon Guard Duty

Now that I have mentioned Amazon SageMaker, let me talk about the challenges we were facing in our fraud detection Pipeline.

Currently, we are doing a lot of preprocessing using a fleet of AWS Lambda functions and the models are trained on physical laptops. While the job is getting done, we struggled to handle large datasets due to the machines’ limited hardware. We were having problems optimizing our model building and data preprocessing pipelines for fraud detection as well as risk assessment.

In one of my conversations with AWS, they routed me to CloudThat and I should really appreciate the time and effort the CloudThat team has put into coming up with the solution and the help they have offered in implementing the same. CloudThat’s team has efficiently leveraged advanced Machine Learning services that are provided by AWS and is up-to-date about the new features that have been recently launched.

ML OPS AWS Diagram    

Let us get down to the technical stuff now and understand ML Ops:

Machine learning operations are a set of procedures for enhancing the quality, simplifying the methodology, and automating the deployment of machine learning (ML) in massive production systems. In contrast to DevOps, MLOps is focused on machine learning ML systems. Data Scientists and developers can automate and standardize procedures throughout the Machine Learning lifecycle by using the machine learning operations (MLOps) tools that Amazon SageMaker offers. To increase the productivity of data scientists and ML engineers while maintaining model performance in production, you can quickly build, evaluate, debug, deploy, and manage ML models at scale using SageMaker MLOps tools.

Amazon SageMaker is a machine learning service that is completely managed by AWS. Data scientists and developers can use SageMaker to construct and train machine learning models fast and efficiently, then deploy them directly into a production-ready hosted environment. The advanced capabilities SageMaker had in books, drove us to choose the service.

SageMaker Capabilities     

There are several features that we have used from the plethora of capabilities that Amazon Sagemaker has to have on offer. To name a few:

1. Amazon SageMaker Studio: 
   1. A development environment with fully integrated ML capabilities
   2. Enables end-to-end ML lifecycle

2. Amazon SageMaker Model Training
   2. Supports Algorithms like XGBoost
   3. Capabilities of Tweaking Hyperparameters
3. Amazon SageMaker Feature Store
   2. Capture and store the features of your dataset
   3. Capabilities of creating groups
   4. Rest API to ingest the data to the feature store for both Realtime(online) and batch data(offline)
4. Amazon SageMaker Clarify
   2. As the name suggests clarifies/explains the model
   3. Works at all the levels: data preparation, model training and model inference
5. Amazon SageMaker Pipelines
   2. Build and manage end-to-end ML pipeline
   3. Supports Model Monitoring
6. Amazon SageMaker Model Deployment
   2. I Think the name is enough for the service
   3. Just one addition, AutoScaling is set up by itself for deployments
7. Amazon SageMaker Processing
   2. Processes our data with no headache of infrastructure management
   3. We can also bring our own custom environment as docker images
8. Amazon SageMaker Model Monitor
   2. SageMaker model monitoring helps to keep track of the models for a certain threshold.
   3. It also helps us to check the biasing/performance for a particular model.
   4. It helps to alert the user via publishing its metrics into the CloudWatch Alarm.

Further Exploration on Amazon SageMaker

I feel we have barely made use of the phenominal service and we should be striving to make use of some of its features like:

• Data Wrangler – With SageMaker Data Wrangler, we can simplify the process of data preparation and feature engineering and complete each step of the data preparation workflow from a single visual interface. This typically includes data selection, cleansing, exploration, and visualization

• AutoPilot – Next talking about Amazon SageMaker Autopilot, it eliminates the heavy lifting of building ML models. We can simply provide a tabular dataset and select the target column to predict, and it will automatically explore different solutions to find the best model.

• GroundTruth – With GroundTruth, Amazon SageMaker enables to identify raw data, such as images, text files, and videos, add informative labels and generate labeled synthetic data to create high-quality training datasets for your machine learning (ML) models.

• Geospatial ML – Amazon SageMaker supports geospatial machine learning (ML) capabilities, allowing data scientists and ML engineers to easily build, train, and deploy ML models using geospatial data.

• Jumpstart – With Amazon SageMaker JumpStart, we can access built-in algorithms with pre-trained models from model hubs, pretrained foundation models to help us perform tasks such as article summarization and image generation, and prebuilt solutions to solve common use case.

I am sure Bhavesh and CloudThat will be excited to extend their hand in exploring these with my team and I.

Key Take Aways

CloudThat did a fantastic job with the task at hand. The implementation has been already discussed in depth by Bhavesh, however, let me talk to you about my personal key takeaways from the solution.

1. Ease of solution suggestions and implementation

Let me give an example. As we are processing a lot of data from credit score providers, not every time do we receive the data in a readable format. We had one such hurdle in this project, and my team highlighted that this would not be possible. Kudos to the CloudThat team for integrating AWS Lambda and Amazon SageMaker so seamlessly that the team did not even have to lift a finger to modify/transform the response.

2. Cloud Native is the way forward than Hybrid

In the previous system, I had to rely on individuals to build and deploy models while the preprocessing was done on AWS. With the solution implemented, there has been no such dependency. Individuals can now focus on monitoring the model performance rather than worrying about data drift.

3. Development to deployment is faster

As Bhavesh Mentioned, deployments are now 50% faster than in the older system. Due to this, the feature release can be quicker with new features released every week as compared to the older system which took over months.

4. Increased observability

In the previous system, I did not have any view of what is happening with my ML ecosystem. With the current system, I can check how a particular model is performing at any given point in time.

Overall, I would term this project a successful implementation of the MLOps pipeline for Risk Assessment and Fraud Detection by synergizing multiple AWS Services.

To encapsulate, with the help of Amazon Sagemaker and CloudThat’s expertise in ML Workflows, we are able to reduce not only the manual dependency but also the purchase of high-configuration laptops along with an increase in the speed from data to delivery.

Further Integrations planned for scaling up the business

Like I mentioned, apart from the exploration on Amazon Sagemaker, during our mid-stage growth, we created a solution that greatly aided the scaling of our business verticals. We noticed that there was a chance to offer this solution to other businesses in the loan industry even though it has grown in maturity in terms of how it addresses the issue statement. In India, the demand-supply gap in capital requirements is sufficiently great that a solution to close it can be provided as a separate service. We are working on this technology mindset and hope to replicate our success with others.

Thank you for your time Back to Bhavesh