Galudy

Galudy Application Infrastructure

Inspiration

The idea was inspired by my company's (Scout24) summer festival party. I wore a Nigerian outfit and people complimented the outfit and asked where they could get it. I knew it would be hard for any none Nigerian to know the names of these kinds of attires so I decided to build some image recognition tool to address this.

What it does

It simply allows a user to do a visual search for any African attire he/she likes, then the solution responds with similar attires. User can then click on any outfit of interest and proceed to buy from the seller. We also added a discovery feature where users can like or dislike pictures for instant recommendations - just like Amazon.com/scout.

How we built it

We built the prototype in 3 days. We didn't want to have anything to do with model training likewise managing any servers. Since, I was already familiar with AWS ecosystem, I looked at the available AWS solutions that would fit the architecture sketch on my notepad and went on to build. We also ensure all infrastructure was deployed via Cloudformation using AWS Serverless Application Model.

The application was developed in two parts

The background job which scrapes external data and indexes AWS Rekognition image detection labels into elasticsearch for querying
The API which is served via AWS API Gateway and AWS Lambda
The user facing which is a static website hosted on AWS S3 with which user interacts with the API
We make use of AWS Comprehend to extract keywords from user search request and before querying our search.
We use AWS Translate to localize product details when a user views it

Challenges we ran into

Main challenges were:

Nigerian outfits are very peculiar in design and image detection, hence a bit on non matching responses from AWS Rekognition. However, this didn't stop the solution from recommending similar items. To our surprise, it performed better than we expected.
Although we built for African outfits, we didn't think of a wider reach in terms of more popular clothe types
Getting more users on board

Accomplishments that we're proud of

The fact that we were able to come up with a working prototype in 3 days
We ran some ads and we had over 3k users test the solution out
The fact that we were able to scale with usage with zero upfront cost
We have also expanded to western outfits
We have both web and mobile versions of the application. Our mobile app (Android APK file) is available for download here

What we learned

Build serverless first
Have a repeatable infrastructure and architecture
Image recognition does not have to start with knowing deep learning
Leverage cloud economies of scale provided by using managed services from AWS

What's next for Galudy

Making it available for other outfits in other regions (/)
Turn in to the most popular visual shopping platform
Create a business model around the solution
Raise funding to expand
Train custom models (tensorflow or MxNet or Gluon on AWS Sagemaker) to improve recommendations.