Erik Quintanilla

Selected work

• BlenderBench

  Creator

  A benchmark for agents working in Blender, from modeling and lighting to rigging, animation, and simulation.

• SWE-Marathon

  Co-author

  A benchmark for evaluating agents on ultra-long-horizon software engineering work.

• computer-1

  Creator

  A model-agnostic computer-use harness inside Harbor.

• Workbench CLI

  A TUI for managing coding agents in tmux.

Background

Before Refresh, I was lead MLE at LynkAI, working on large-scale voice data and models, and an early engineer at Superdial, working on IVR navigation and evals around medical phone calls using GPT 3.5 Turbo.

Earlier, I worked on large-scale data at Capital One and scraping at Amazon, where I submitted a patent in 2021 on automatically generating scrapers.

My research focused on GAN-based user preference modeling and cross-view synthesis (ACM-MM ’20, ECCV ’20).

College projects & competitions

In college, I worked on financial modeling through the Citadel Aperture program and competed on Numerai (2× Gold, 99th percentile; 5× Silver in Numerai Fund’s Weekly AI/ML Competition). I also worked on adversarial synthetic data generation extensively for the Numerai competition.

I won hackathons hosted at UChicago, Vanderbilt, and Amazon.

Writing

• You Can’t Round-Trip Audio Through Log-Mel

  Why audio to mel to audio is lossy, why vocoders exist, and when the gap matters.

• Recoverable Before Readable

  Reading transformer residual streams as trajectories to catch jailbreaks early, audit interpretability probes, and watch decisions form (Qwen3.5-2B).