Jon Maravilla

Jon Maravilla

Undergraduate researcher, Efficient AI Lab at Oakland University

  • Sustainable AI
  • Efficient agentic systems

About

Hi there! My name is Jon, I'm a computer science sophomore at Oakland University and a research assistant in the Efficient AI Lab, advised by Dr. Kaiqi Zhao. My research interests include sustainable AI and efficient agentic systems. Recently, I've focused on modeling the runtime and energy consumption of diffusion models.

Outside of research, I previously competed for Team USA in figure skating, earning a junior national medal and reaching a top-60 world ranking. I still currently coach and choreograph part-time.

Recent updates

  1. Nov 2026

    Paper accepted to the Interpreting Agent Behavior (IAB) Workshop at NeurIPS 2026 (First-author).

  2. Oct 2026

    Won Hacktoberfest x GrizzHacks at Oakland University.

  3. Sep 2026

    Selected for the NSF S-STEM Scholarship program.

  4. Jul 2026

    Paper accepted to the Trustworthy AI4Good Workshop at ICML 2026 (First-author).

  5. Mar 2026

    Paper accepted to IJCNN 2026 (Co-authored).

Publications

ROTOM architecture: a deterministic game-family router sends each GLEE game instance to a bargaining, negotiation or persuasion specialist, which produces the action.

ROTOM: Task-Specialized Inference and Multi-Agent Orchestration in Economic Games

ROTOM: Task-Specialized Inference and Multi-Agent Orchestration in Economic Games

Jon Maravilla, Juan Carlos Maravilla, Kaiqi Zhao

IAB Workshop at NeurIPS 2026 — GLEE Competition

Explores task-specialized inference and multi-agent orchestration across bargaining, negotiation, and persuasion tasks.

DiTCarbon pipeline: DiT architecture and generation configs feed parameter, hardware-efficiency and FLOP models, which drive operational and embodied carbon, summed into total CO2eq.

DiTCarbon: Predictive Carbon Footprint Estimation for Diffusion Transformer Inference

DiTCarbon: Predictive Carbon Footprint Estimation for Diffusion Transformer Inference

Jon Maravilla, Justin Kur, Kaiqi Zhao

Trustworthy AI4Good Workshop, ICML 2026

Predicts the energy and carbon cost of diffusion-transformer inference, cutting energy-prediction error 36–59% vs. baselines.

Attention heatmaps comparing the teacher model, Trivial Augment and standard data augmentation.

Data-Augmented Quantization-Aware Knowledge Distillation

Data-Augmented Quantization-Aware Knowledge Distillation

Justin Kur, Jon Maravilla, Kaiqi Zhao

International Joint Conference on Neural Networks (IJCNN), IEEE WCCI 2026

Uses data augmentation to strengthen quantization-aware distillation, raising Top-1 accuracy by up to 10% for sub-8-bit CNNs.

Experience

  1. Undergraduate Research Assistant

    Efficient AI Lab, Oakland University. Advised by Dr. Kaiqi Zhao.

    2025 – Present

  2. Co-founder, Ysari Digital

    Websites and digital marketing for small businesses

    2023 – 2025

Contact

The best way to reach me is email.

imaravilla@oakland.edu