Technical Directory for a wanderplay in Los Angeles, CA. Responsible for all technical aspects of the show, including web development, raspberry pi integration, and custom SIP media phone menus.

Studied approaches to AI image generation based on musical and textual inputs. First approach is lyric-based, producing image-generation prompts using LLMs via a synthesis of textual analysis and summarization of lyrics. Second approach fine tunes a MIR model to describe images which correspond to provided musical snippets. Final approach fine-tunes a Stable Diffusion XL model to directly produce images from audio embeddings provided by AST. Results show that the lyric-based approach is the most successful, while other approaches are highly data-constrained.

Semester project implementing core concepts from AlphaGo and AlphaGoZero to train a Go playing agent. Major accomplishment was using ZeroMQ and small networks to colocate multiple inference instances on a single consumer-grade GPU.

Semester project studying the applicability of Encodec towards producing general purpose audio embeddings. Our approach encodes AudioSet using Encodec and then trains MobileNetV3 to classify the original AudioSet tags. We evaluate the performance using linear probing and find that while the approach underperforms in specific tasks, it performs surprisingly consistently across all tasks.

Python library for interacting with Fabman.io API for maker space access control. Developed to transition TIW maker space from homespun solution to Fabman.io.

Using LLM to synthesis Faust code to build digital synthesizers. Based loosely on AlphaCode, a LLM produces novel Faust code to build new audio synthesizer architectures.

Final Project for Robot Learning graduate course. Comparison of different torch data strategies resulting in 33% more efficient data loading for large multi-modal datasets.

Final project for undergraduate ML intro course. Used machine learning to dig into the Gaia Data Release version 3 to identify celestial objects.

Using few-shot learning and Pattern-Exploitative Training to identify spam and sentiment at the heart of the 2022 crypto boom and bust.

Using python and numpy to simulate gates for quantum computing calculations.

Web Portal application for Texas Inventionworks, the makerspace located in the heart of the Engineering Education Research building at UT Austin.

Python library for parsing SysEx information from the Elektron Digitone. No longer maintained to parse newer firmware.