CHRISTOPHER AKATSUKA

Breaking Up The Monolith

Summary

As software lead, I worked with four software engineers to build towards a micro frontend target state that reduced deployment times from six hours to 20 minutes.

Technical Details

A Capital One web application called Dealer Navigator allows auto dealers to finance their customers' auto loans. The document upload section, one of the core services offered, receives hundreds of thousands of visitors per day and generates over a billion dollars in revenue every year. Unfortunately, the application became a monolith over the years and desperately needed a rework.

I led the effort in building an Angular micro frontend with a dedicated backend for frontend for the document upload section. In addition to a normal developer workload, I reviewed pull requests, guided architecture discussions, and provided one-on-one mentorship sessions. By the end of the project, we achieved our goal of having a modular and refactored codebase for our frontend and backend.

Beyond dockerization and migrating code out of the monolith, I made many technical improvements. I hosted the dockerized containers on an Amazon Web Services ECS cluster, automated the run of perf and functional tests before each release, and added route guards to prevent URL tampering. I also moved all application logic to the BFF layer, updated deprecated dependencies, and improved test coverage to surpass company standards.

Reimagining The Customer Experience

Summary

I rebuilt a business-critical Angular web app with loan funding requirements, document classification, and a PDFJS document previewer, resulting in a 15% increase in dealerships leveraging the web app over mail and fax.

Technical Details

I integrated NGRX for maintaining component state across different pages, built collapsible banners for previewing uploaded documents, and styled tables to clearly communicate dealer funding requirements. In order to leverage all of PDFJS's features, I also built a Lambda function on Amazon Web Services that created a PDF copy of every image uploaded to S3.

Through this project, I learned how to build responsive pages that could work across all modern browsers. I had to learn about breakpoints, polyfills, and the wonders of caniuse.com.

Clearing Up The Confusion

Summary

I architected and developed a Java Spring Boot REST API that provides application-specific funding requirements, reducing the number of applications with missing documents by 55%.

Technical Details

I leveraged the Builder design pattern to program business logic into readable and maintainable classes. This allowed me to modularize the logic for including universal requirements, special requirements due to application parameters, and custom requirements from funders.

Dealers responded to these changes with overwhelmingly positive feedback, and the number of "clean deals" (deals that provide all required documentation on the first try) increased drastically.

Classifying And Extracting Documents

Summary

I replaced a manual document classification process with an event-driven Python microservice pipeline on Amazon Web Services that classifies 1 million documents per day in less than 10 seconds per page, saving $6 million annually.

Technical Details

For the pipeline, I built a dockerized image preprocessing Python microservice. The tool leverages OpenCV to grayscale images, convert files to JPEG, and perform content isolation transformations to increase accuracy. I deployed the container on an Amazon Web Services ECS cluster.

When an upload occurs, the orchestrator places a message, which my preprocessing service consumes, in an SQS queue. Upon completion, my service saves the processed image in an S3 bucket and places a message for the next microservice in the pipeline to consume. In the event of failure, the service retries up to three times before placing a message in a dead letter queue for manual review.

I also built an Angular 4 GUI to help retrain the classification model. The GUI allows users to view uploaded images next to their classifications. If a user identifies a misclassification, he or she can select the correct label from a dropdown. Reclassification data feeds directly into retraining data sets, which help to improve classification accuracy.