MSc Computer Vision @ MBZUAI

Juan Nicolas Sepulveda Arias

AI safety, security, and computer vision researcher focused on robustness evaluation, watermarking, and generative model security.

Juan Nicolas Sepulveda Arias at BSIDES Colombia 2024
BSIDES Colombia 2024 Generative models for cyberthreat detection
2026 Research internship candidate
95% Malware benchmark accuracy
0.65 mAP50-95 after YOLOv8 tuning
Top CGPA in undergraduate cohort

Research Focus

Safety-centered work on generative models and visual systems.

I work on AI safety and fairness for generative models, with emphasis on adversarial robustness, red-teaming, and content authenticity.

Generative Model Safety

Evaluation of failure modes, fairness metrics, and security risks in modern AI systems.

AI safety Fairness Evaluation

Robustness & Red-Teaming

Jailbreak robustness, adversarial testing, and failure analysis for LLM and vision systems.

Red-teaming Jailbreaks Security

Watermarking & Authenticity

Content provenance and watermarking methods for diffusion models, LLMs, and generated media.

Watermarking Diffusion models LLMs

Featured Work

Projects presented as evidence, not just entries.

The strongest work is grouped around outcomes, methods, and technical signals visitors can scan quickly.

Malware generation + detection with interpretability

Using GANs to Detect Cyberthreats

Trained a GAN on VirusShare to detect and generate obfuscated and non-obfuscated binary malware samples, with interpretability and reverse-engineering analysis.

Accuracy
92% to 95%
Methods
GANs, GradCAM, hooks
Domain
Cybersecurity

Weather-robust detection with YOLOv8

Traffic Signs Recognition Systems

Fine-tuned YOLOv8 on GTSRB with simulated fog, night, and rain, then tuned the final layer to improve detection robustness.

mAP50-95
0.39 to 0.65
Stack
YOLOv8, GTSRB

Experience

A compact timeline of research, engineering, and leadership.

MSc in Computer Vision, MBZUAI

Fully funded scholarship with coursework in generative AI safety, ML security, and visual computing.

Graduate Researcher, MBZUAI

Research under Nils Lukas on AI fairness metrics and watermarking methods for generative models.

RLHF/SFT Python Coder, Turing

Produced code-writing examples and rubric-based evaluations for model failure analysis.

Software Engineering Intern, DevSavant

Built and deployed a Django and OpenAI API MVP for assisted autobiography writing.

IEEE Computer Society Student Chapter President

Organized competitions, executive AI workshops, and the Code The Future hackathon.

Technical Toolkit

Grouped for scanning across research and implementation.

ML / AI

PyTorch HuggingFace Lightning FastAI GradCAM Failure analysis

Development

Django FastAPI Git / GitHub Linux GCP

Languages

Spanish native English C1 Portuguese B2

Contact

Open to 2026 AI safety, security, and generative model research internships.