AI Safety & Interpretability
ML Engineer · New York, NY
Machine learning engineer transitioning to AI safety and alignment research. My primary interest is in interpretability and its applications: understanding the inner workings of neural networks and building robust techniques to make AI systems safer. Originally from Santa Rosa de Osos, Colombia 🇨🇴, and currently based in New York City 🗽.
If you’d like to share your thoughts or just say hello, you can reach me at [email protected].
Course & research projects
Technical AI Safety Project: Round 2
Jul. 2026 – Aug. 2026
Five-week research sprint
Project name: Sandbagging Detection with Natural Language Autoencoders: A Reliability Audit on a Password-Locked Organism
Write-up and repository coming soon
Technical AI Safety Project: Round 1
Apr. 2026 – May 2026
Five-week research sprint
Project name: What synthetic-document fine-tuning transmits, and what it doesn’t: a layered dissociation in evaluation-aware behavior on Qwen2.5-7B
Technical AI Safety
Mar. 2026 – Apr. 2026
Six-week course
Completed this program covering AI safety threat models, evaluating current safety techniques, and identifying high-impact intervention points.
Self-study
Alignment Research Engineer Accelerator
Feb. 2026 – Apr. 2026
Self-studied the curriculum, focusing on Chapter 0 (deep learning fundamentals) and Chapter 1 (interpretability). Below is my companion repository with current progress.
M.S. in Computer Science
Sept. 2020 – May 2022
Focus on Machine Learning.
B.E. in Mechanical Engineering
Sept. 2015 – Jun. 2019
Minor in Computer Science.
American Express
New York, NY
Machine Learning Engineer
Jun. 2022 – Present
Building and shipping customer-facing GenAI products.
American Express
New York, NY
Edward P. Gilligan Award for Innovation
2026
Recognizes colleagues for their innovation, partnership, and commitment to delivering meaningful impact.
Inventor Award
2025
Recognizes colleagues for their contributions to the company’s patent portfolio.
Centurion Award
2025
Recognizes approximately the top 1.5 percent of colleagues for their exceptional performance, leadership, and impact.
June 8, 2026
Synthetic-document fine-tuning on Qwen2.5-7B installs the corpus belief at the recognition layer, not in generation, and the only task-behavior effect attaches to the wrong cue.
