Juan Alvarado

Software Engineering Graduate Student at Carnegie Mellon.
Google TechX & NSF REU Fellow.

About Me

I'm a disciplined and determined first-generation college student, now a graduate student at Carnegie Mellon University focused on software engineering and machine learning.

As an NSF REU fellow at the University of Texas at Dallas, I fine-tuned LLMs and researched generative vision-language models. During an internship with the City of Coral Gables' GIS Department, I set up locally hosted speech recognition tools such as OpenAI's Whisper. I'm also a Google Tech Exchange 2025 alumnus and have volunteered at a local community center for nearly five years.

Currently seeking: SWE & ML roles.

Projects

First page of the paper Fine-tuning TinyLLaVA for Multimodal Sarcasm Detection on MMSD 2.0

Fine-Tuning TinyLLaVA for Sarcasm Detection

Fine-tuned TinyLLaVA, a 3.1B-parameter vision-language model (Phi-2 + SigLIP), with LoRA to detect sarcasm in image and text posts from the full MMSD 2.0 dataset, evaluated with accuracy, F1 and MCC. NSF REU research at UT Dallas, Summer 2025. Unpublished manuscript.

PyTorch LoRA Vision-Language Models Hugging Face

Raspberry Pi 4 wired to a USB microphone and a camera module on a desk

[ IN PROGRESS ]

Raspberry Pi Desk Assistant

Work in progress: a team project with three classmates for a data science course at Carnegie Mellon. A Raspberry Pi 4 voice assistant that wakes on a custom wake word, recognized enrolled teammates on-device (OpenCV YuNet + SFace embeddings), transcribes questions with Whisper, answer questions with an LLM and replies aloud with Piper, all run by an asyncio state machine.

Python Raspberry Pi OpenCV Whisper LLM APIs

Workout Activity App

Built a full-stack fitness tracker web app with a 4-person team in Google's Tech Exchange program (Spring 2025), using Streamlit and Google Cloud BigQuery for the SQL database.

Python Streamlit Google Cloud BigQuery SQL

juan-alvarado.com

This site, hand-built with no framework: a real-time Three.js background, a PDF.js résumé viewer and a custom photo gallery. Performance work cut the desktop load time on a 4G connection from about 7 seconds to about 2.

JavaScript Three.js HTML CSS Vercel

StudyBudy App

Led a 4-person team at ShellHacks 2024 to build a full-stack web app that turns uploaded PDF notes into quizzes using ChatGPT.

JavaScript Node.js Express MongoDB

The finished rig with RGB-lit graphics cards and tidied cables

Ethereum Mining Rig

With friends, I built and ran two GPU mining rigs (five Nvidia RTX 3080-series cards and five overclocked AMD GPUs) on HiveOS, a Linux-based mining OS, until Ethereum moved from proof of work to proof of stake in 2022.

HiveOS Linux GPU Overclocking