Cloud Tenant Networking API:
End-point centric cloud deployments

With cloud deployments and network virtualizations among popular providers (such as AWS, Azure, etc) becoming more complicated and expensive, we implemented a solution for users to simplify the creation and operation of these virtual networks by creating a declarative API.

This wrapper API library abstracts away the complicated specifications such as GPU size, elastic IP addresses, load balancers, etc. so that users can just call the API endpoints to create a virtual machine in just a few minutes!

For more details, read the full paper here

This project was a great learning experience to deep-dive into computer networking and how virtual machines are provisioned by calling these providers.

Using the python SDKs for AWS and Azure, we were able to strip down the complexity and logic of creating load balancers, requesting elastic ip addresses, provisioning virtual machines, setting permissions, and wrap them in neat endpoints so that the user can easily call these methods to create their own virtual machines on any cloud platform! This allows any user to essentially create their own virtual network without having to worry about all the complicated specs like manually creating their own Virtual Private Clouds, assigning subnets and allocating ip addresses, etc.

This is particularly useful for those who want to quickly create a virtual machine and get it up and running in just a few minutes without wasting time switching or onboarding between multiple cloud platform accounts. Go ahead and try it out yourself: Github repo

Memories:
Image-to-emotion classifier

During my first semester in Launchpad (a UC Berkeley machine learning community for students), our team created a web app where users can upload pictures from their photo album and generate ambient sounds and songs from the image’s scenery and emotion.

Check out our website here!

Our software uses pre-trained models such as Resnet and Soundnet and a multimodal architecture to deconstruct both images and audio.

We trained datasets consisting of over thousands of images and audio samples by building off the Image Resnet model and Sound MFCC model to output an emotion vector for each song given an image. Then we compared the vector features by calculating their cross-entropy and selected the top 3 audio clips with the greatest similarities with the sample image.

The web app was created in React and hosted on Firebase.