← back to the network

// model_card

marcel-mateos-salles

The static, printable readout of the interactive network at marcelmatsal.com.

---
model: marcel-mateos-salles
architecture: 9-layer experience stack · 2 probe-able unit layers
base_model: brown-university/cs-econ (magna cum laude, honors)
finetuned_for: [backend engineering, ML research, interpretability]
release: v2026.05
license: open-to-opportunities-1.0
contact: marcel_mateos_salles@alumni.brown.edu
---

Model description

Backend software engineer and ML researcher. Recent graduate of Brown University (Computer Science-Economics), member of the GalilAI Group with Prof. Randall Balestriero, and incoming Software Engineer at Pinterest. Research interests: LLMs, model interpretability, and self-supervised learning — with a particular focus on spurious correlations under parameter-efficient finetuning.

Architecture

One unit per role, stacked in temporal order. t-0 is the most recent layer; activations reflect how central each unit is to the current model.

layerunitsourceperiodact
t-0Pinterest SWEPinterestIncoming July 20260.78
t-1Head TA · CLCSCI1460: Computational LinguisticsDec 2025 - May 20260.60
t-2Pinterest InternPinterestSummer 20250.74
t-3TA · Deep LearningCSCI 1470: Deep LearningDec 2024 - May 20250.62
t-4GalilAI ResearchGalilAI Group @ Brown UniversityOct 2024 - Present0.96
t-5Dexcom InternDexcomJun 2024 - Aug 20240.70
t-6TA · Big DataEcon 1000: Using Big Data to Solve Social and Economic ProblemsSep 2023 - Dec 20230.55
t-7SAVE LabSAVE Lab @ UTSAMay 2023 - Sep 20230.82
t-8Brown · CS-EconBrown UniversitySep 2022 - May 20260.90

Training data

  • Brown University — B.S. Computer Science–Economics, Magna Cum Laude, Honors. GPA 4.0/4.0.
  • Reinforced by teaching: Deep Learning (CSCI 1470), Computational Linguistics (CSCI 1460, Head TA), and Using Big Data to Solve Social and Economic Problems (Econ 1000).
  • Fine-tuned in production at Pinterest (GenAI tooling, observability) and Dexcom (mobile architecture).

Evaluation

LoRA Users Beware: A Few Spurious Tokens Can Manipulate Your Finetuned Model

Marcel Mateos Salles, Praney Goyal, Pradyut Sekhsaria, Hai Huang, Randall Balestriero

NeurIPS 2025 workshop · poster · ICLR 2026 workshop · oral · arXiv

When Efficiency Enables Shortcuts: Studying Spurious Correlations Under LoRA Finetuning

Marcel Mateos Salles

Brown University honors thesis · thesis · PDF

Feature dictionary

Java f#3650Python f#3324TypeScript f#1221JavaScript f#0429R f#0082CSS f#1507HTML f#4011Pytorch f#2931TensorFlow f#2375Spark f#0701JavaFX f#3732React f#2399Next.js f#3588Tailwind CSS f#1595GraphQL f#1993Pandas f#3577Numpy f#4079Scikit-learn f#3252AWS f#3805Flask f#4051HuggingFace f#0498PEFT f#2435

Intended use

Backend systems and ML research teams that want an engineer who reads the papers and ships the infrastructure. Direct use: marcel_mateos_salles@alumni.brown.edu. See also GitHub, Google Scholar, LinkedIn, and the resume (PDF).

Limitations

  • Performance degrades measurably when the coffee supply is ablated.
  • Addicted to fitness and spending time outside. Currently training for the Seawheeze Halfmarathon.
  • Will occasionally overfit to a random fixation.

Citation

@misc{mateossalles2026,
  author = {Mateos Salles, Marcel},
  title  = {marcel-mateos-salles: an interpretable portfolio},
  year   = {2026},
  url    = {https://www.marcelmatsal.com}
}

// generated from the same data that powers the network