Joseph Ellis

Joseph Ellis

AI Software Engineer at BuildyLos Angeles

I build AI products. The model is usually the easy half.

I have a background in machine learning and data science. My own projects explore language-model agents, computer vision, and making data easier to work with.

Mostly TypeScript, Python, React, and PostgreSQL.

Selected projects

A research framework for language-model agents playing a social deduction game. I developed game logic, agent reasoning modules, and an evaluation pipeline to compare how agents use incomplete information.

Compared agent behavior across 499 games and five configurations.

ChatDB

A natural-language interface to an NBA database. I built the data preparation pipeline, SQL generation, and a Streamlit interface that shows the query alongside its explanation.

Separate paths for reading, exploring, and modifying data.

NBAR

An NBA data pipeline packaged for use in R. I connected scheduled box-score collection, data cleaning, and package distribution so the analysis starts with reusable data.

An installable package with scheduled collection and testthat tests.

A computer-vision study of whether video content can predict engagement. The project combines a custom dataset, two definitions of virality, and fine-tuning of a pretrained video model.

The evaluation exposed prediction collapse rather than a reliable predictor.

More projects
  • PLG SaaS Marketing Mix Modeling

    Bayesian MMM in R with adstock decay, Hill saturation, and hierarchical priors, sampled via brms and Stan.

    Code
  • DrugAI-CVAE

    Autoregressive conditional VAE generating novel SMILES sequences conditioned on target proteins, with two faculty collaborators.

    Code
  • Notion Work Scheduler

    Time tracking synced to a Notion database, with local storage fallback so it works before Notion is configured.

    Code
  • Fire Segmentation

    U-Net semantic segmentation with Grad-CAM interpretability, to check what the model attends to rather than trusting the metric.

    Code
  • Pest Patrol

    Real-time detection on an NVIDIA Jetson driving a servo over GPIO, threaded so actuation never blocks the inference loop. A prototype.

    Code
  • ASL Classification

    CNN classifier on the ASL alphabet, with model code and analysis code cleanly separated over a shared preprocessing pipeline.

    Code
  • Psyche Evaluation

    Python and MySQL application administering the Big Five and NPI inventories, with scoring, aggregates, and export.

    Code
  • Money Wicket

    Cricket transfer scouting via logistic regression with Ridge and Lasso regularization.

    Code
  • Soccer Linear & Logistic Analysis

    A self-sourced Premier League dataset run through linear and logistic models in pandas.

    Code

Experience

Buildy

AI Software Engineer

Build customer-facing products at an AI business platform: social publishing, paid advertising, AI media generation, and course creation.

  • Built a course-creation product covering lesson authoring, pricing and offers, publishing, narration, and an AI tutor, with payments going to the creator’s own Stripe account.
  • Built publishing integrations for Facebook, LinkedIn, X, and TikTok, and extended them to multiple accounts per platform.
  • Developed Google Ads connectivity, conversion tracking, and campaign creation for Search and Demand Gen.
  • Shipped image and video generation across several model providers, with credit metering and publishing to social platforms.
  • Built context management for the chat agent so long-running projects keep their history.

Started on marketing tools and ended up building a course product from nothing to something creators can sell.

PushButton.AI

Machine Learning Engineer

Built the LLM backend for an automated course-generation service: content pipelines, async job processing, and model evaluation.

Where I started treating model choice as an engineering decision rather than a preference.

Ovrsee

Co-Founder, Backend Engineer

Built the backend for a TikTok creator analytics product, including account connection, scheduled data collection, and reporting with Firebase and Google Cloud Functions.

The same TikTok auth and token-refresh problem I would meet again eighteen months later.

Cabrillo Marine Aquarium

Data Science Intern

Organized animal welfare data and created visualizations for aquarium exhibits. Set up repeatable reporting so aquarists could continue using the work after the internship.

First time building something whose main requirement was that someone else could keep it running.

Background

University of Southern California

MS, Applied Data Science

Viterbi School of Engineering · Los Angeles, CA

2024 – 2026

Chapman University

BS, Data Analytics

Fowler School of Engineering · Minor in Business Administration · Orange, CA

2020 – 2024

Honors

  • Magna Cum LaudeChapman University
  • Provost's List — 7 semestersHonors standing every semester enrolled
  • Excellence in Computing StudiesAssociation for Computing Machinery