About Me

I'm a senior majoring in Computer Science and Music and minoring in Political Economy at the University of California, Berkeley. I have a deep love for learning, engaging with complex and new ideas, and problem solving in a variety of fields. I'm interested in software engineering and design, graphics, deep learning, and utilizing software and technology to tackle social and industry problems. I'm hoping to find a meaningful role to continue to grow and learn to solve new, exciting problems!

Additionally, I'm pursuing a personal project in the form of writing and publishing music under the artist profile Sam Ellgass while continuing to market and build my personal brand (click on the icons above to take a listen!). My ongoing personal projects are the Royal Ghost Clash Royale AI agents and the BerkeleyMon Pokemon-style game clone, both written from scratch in Python.

Featured Projects

BerkeleyMon: Pokemon AI Agents (Summer 2022)


(Pictured) A randomized wild encounter against an AI agent playing as Infernape.

• Goal: Create a pseudorandom Pokemon game with functionally unbeatable Artificial Intelligence that's able to select an action in real time.
• Implemented all 200+ Generation 4 Pokemon and 120+ Pokemon moves.
• Created Pokemon game logic from scratch, including switching, items, catching logic, leveling, evolution, precedence, and inherent randomness.
• Implemented a Monte Carlo Tree Search method using "Laplacian Mini-Rollout" (described further in the paper hyperlinked below).

Royal Ghost: Clash Royale AI Agent (Spring 2022)


(Pictured) The gameboard in both visual representations.

• Goal: Train an Artificial Intelligence agent to play Clash Royale rationally and effectively.
• Findings: Using reinforcement learning (Q-learning), the agent was able to learn when to use key cards in just over 5,000 games.
• The gameplay was simplified by making all turns occur within one update pass every second, scaled to be accurate to the real game, and this allowed for training with full games at 100x speed.
• The Q-Learner agent learned by playing successive games against an equivalent agent with older weights, updating the weight of each Q-value after each turn concluded.
• The Q-Learner agent was biased toward cheaper cards with more easily verifiable Q-values, and learned how to properly counter different types of cards (air units, tank units) by playing against itself.
• Next, a neural network based agent will be trained to play against the Q-learner, and the agents will be playable opponents for a human user.

CS184 Graphics Portfolio (Spring 2022)


(Pictured) A cornell box + spheres rendering with our ray tracing engine.

• Overview: A portfolio of graphics projects completed in C++ for CS184.
• Rasterizer - a rasterizer for simple polygons, rendering SVG images to the screen pixel by pixel.
• MeshEdit - a triangle mesh editor engine, with the ability to edit CODA files and perform Loop Subdivision to upsample 3d meshes.
• PathTracer - a ray tracing engine for global illumination, including multi-bounce illumination, accurate tracing of trajectories, and importance sampling for efficiency.
• ClothSim - a spring-based physical simulator for cloths and 3d objects, with object collision, Blinn-Phong shading, environment map shading, and the ability to use custom textures.

Spotify Playlist Matcher (Fall 2021)


(Pictured) Adding my song 'THE DAREDEVIL' to a Playlist on Spotify.

• Goal: Given a user login and any song, match the song to the user's most fitting playlists.
• Findings: Currently, using Random Forest multi-class classification, this model is able to determine what playlist a song is most like, with slight bias toward very large and very small playlists.
• Using Python (Flask, Spotipy), built a web application to allow user OAuth to Spotify.
• Created a Python package (Sklearn, Numpy), to build a prediction model from a user's proprietary public and private playlists.
• Combined this package with my flask app to allow a user to search any song and add it to their playlist with the click of a button.

Call Me Ace Internship (Spring 2021)


(Pictured) Distribution of posts with advancing to round 3 as hue & means displayed.

• Goal: create insights from social media data and modeling.
• Findings: reach is less important than frequency of posting (Instagram) in contest success.
• While working as the Product Marketing Intern for Rapper Call Me Ace, I oversaw his data-driven social media presence.
• Here: I worked as a campaign analyst during the BET AmpliFind Competition, in which 50 artists proceeded in the competition through 4 stages based on social media engagement.

LinkedIn Learning Consulting Project (Spring 2021)


(Pictured) Analyzing LinkedIn's data to determine the impacts of our campaigns on viewership.

• Goal: Drive adoption of LinkedIn Learning among Berkeley Students via demographically targeted marketing.
• Findings: Business and Computer Science majors were the most avid LinkedIn users, and most easily adopted LinkedIn Learning as a service.
• Leveraged Python (Pandas, Numpy) to analyze survey results as well as LinkedIn's proprietary data.
• Used these demographic insights to strategically launch marketing campaigns.

SAAS Sexual Education Research Paper (Spring 2021)


(Pictured) The statistical significance of 'Presentation of Sexuality' in State Sex. Ed. Curriculum.

• Research Question: How have State-wide Sexual Education policies of the last 20 years affected the prominence of LGBTQ+ communities across the United States?
• Findings: the representation of sexuality in the classroom has a statistically significant relationship with LGBTQ+ representation in that State, and more informative policy generally correlated with higher LGBTQ+ representation.
• I employed a variety of modeling techniques on US census and sexual education public data, and I worked to assess causes and causality.
• Using Python (Pandas, Numpy, Sklearn), PCA proved especially helpful, as did trying to create hyperfeatures from the largely binary data (i.e. implemented a safe-sex policy or omitted such a policy).