Some Projects I've Worked On
Constantly being updated as I work on more things 😊
Media Planning Tool
This was my first baby while working in Amazon and I love this service so much. I worked on EVERYTHING related to this service which basically serves the purpose of streamlining advertising campaign creations for managed service and also gives recommendations for certain inputs such as budgets for product catalogs and audience targeting segments. I even did a full NAWS migration (moving from internal infrastructure that was dated to the AWS that outsiders use) for the backend of this service which was a major case in my promotion to SDE 2 as we have nearly 100 APIs and I had to manually change every API. I had to create a new AWS ECS service for this migration and even led a design for it which was my first one!
Line Item Budget Recommendations for Amazon Self Service
This was tied to a senior VP goal within my organization. I created a vanilla AWS sagemaker solution that allowed us to get rid of a lot of operational overhead since existing sagemaker wrapper services within Amazon brought in a lot of unnecessary infrastructure in CDK. I worked closely with another team and did away team work in their code base to deliver this goal ASAP. We successfully delivered the goal on time and today we are serving 100% ML recommendations for initial budget allocations for products in Amazon self service that go through an internal tool called HIBOU which optimizes budgets overtime as a campaign is live.
Twitch Contextual Targeting Widget for Managed Service
This was tied to a senior VP goal within the Twitch organization. I invented a simple solution for injecting microfrontends (MFEs) using react-helmet and using custom event listeners to pass data. This was crucial in helping dependent teams deliver for the goal tied to deprecating certain twitch widgets running in managed service advertising consoles in Amazon. Another internal team called OMS was able to use my solution in order to save time and deliver the goal on time.
Don't Think Finance dot com
A simple financial tool for gen alpha and gen z. Brain rotted? No problem, just choose your life style archetype and don't think.
Pocket Golf Coach AI
Used GPT Vision API to create an iOS app that gives you concrete feedback on your golf swing.
Guitar Tab Generator
Made an electron app that takes audio input and maps the audio to a guitar tab. Was part of my buildspace season 4 application.
Neural Network Colorizer
This was the final project for my graduate artificial intelligence course where I created both a 2 layer neural network and a basic K-Means classification agent to recolor the second half of a gray scale image based on the colored first half of it. Both K-Means clustering and the neural network were coded from scratch in order to understand both techniques better as a supplement to the course material.
GoodNews
This was a project for HackDSC in 2020 when the quarantine just started. My friend and I made an iOS app that uses a sentimenet analysis model to filter the good news from the bad news and only show the good news on a feed.
Prioritized Hindsight Experience Replay (PHER) DQN for Difficult Atari Game
This was my final project for my graduate machine learning class in Spring 2021. I was able to conduct a small scale original research project inpsired by the work on HER and Prioritized experience replay in order to convert the Atari game Motezum'as Revenge into a goal based environment for goal based learning for the solution of the sparse reward problem in DRL. This model was able to achieve better rewards than the baseline DQN as was hypothesized and is a good stepping stone for future work.
Focusbyte Avatar Feature
Helped my friend on a pomodoro timer app using React Native. I coded the avatar feature which allows users to customize their in app avatar.
Path Finder Visualizer React
A path finding visualizer for Djikstra's, A*, BFS, and DFS made when I was first learning React.
Robinhood Algorithmic Trading Bot
An algorithmic trading bot which employs the golden cross strategy.
CoRL20: Deep Reinforcement Learning with Population-Coded Spiking Neural Network for Continuous Control
Accepted publication to CoRL20 about DRL for contiuous control using a spiking actor network in some standard actor-critic DRL methods, such as DDPG, TD3, PPO, and SAC. This project showed that the use of spiking neural networks in DRL can allow for better energy efficiency (important for mobile robotics) while not compromising performance.
TMLR22: Deep Reinforcement Learning with Population-Coded Spiking Neural Network for Continuous Control
Accepted publication to TMLR22 about decoding EEG using spiking neural networks (SNNs) on neuromorphic hardware. I mostly trained existing sota DNN methods for decoding EEG data as baselines to compare our SNN method with. Implemented and trained the following models: 3D-CNN, 2D-CNN, CNN-TC, and CNN-LSTM.
Interface for A Neuro-inspired Oculomotor Controller for a Robotic Head Prototype
Coded an interface using Python for a masters student's thesis on a robotic head using spiking neural networks for tracking laser points on a wall.