Juan Nicolas Sepulveda Arias

MSc Student (Computer Vision) @ MBZUAI · AI Safety & Security · Research internship candidate (2026)
Abu Dhabi, UAE · jn.sepulveda55@gmail.com · +971 58 538 1633 · jnsepulveda55.github.io · github.com/JNSepulveda55 · linkedin.com/in/juannicosepulveda · orcid.org/0009-0000-0619-0704

CV PDF

Research focus

I work on AI safety & fairness for generative models, with emphasis on adversarial robustness (jailbreaks / red-teaming) and content authenticity (watermarking for diffusion models and LLMs).

Keywords: AI safety, fairness, red-teaming, jailbreak robustness, watermarking, diffusion models, LLMs, evaluation, interpretability

Highlights

Generative models for malware analysis (VirusShare): improved benchmark accuracy 92% → 95% vs. a prior CNN baseline; added interpretability (GradCAM, hooks) + reverse engineering.
Robust traffic-sign detection: fine-tuned YOLOv8 on GTSRB with weather augmentations; mAP50–95: 0.39 → 0.65 after architecture tuning.
Seeking (2026): research internship in AI safety / generative model security, especially robustness evaluation and watermarking.

Education

MSc in Computer Vision, Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) — Aug 2025–May 2027 (expected)

Fully funded scholarship (UAE government).
Coursework: Fairness & Safety in Generative AI, Machine Learning Security, Deep Learning for Visual Computing, Mathematical Foundations of AI.

BSc in Applied Math and Computer Science, Universidad del Rosario — Jan 2021–Dec 2024

GPA: 4.62/5.0
50% scholarship (top applicant).
Graduated with top CGPA in cohort (recognition of excellence).

Research experience

Graduate Researcher — MBZUAI (Abu Dhabi, UAE)

Aug 2025 – Present

Research under Nils Lukas.
Literature review and synthesis on AI fairness metrics and state-of-the-art watermarking methods for generative models.

Industry experience

RLHF/SFT — Python Coder, Turing (Remote)

Sep 2024 – Dec 2024

Produced ~10/day high-quality SFT/RLHF examples for code-writing and problem-solving tasks.
Performed rubric-based evaluations to identify failure modes (hallucinations, instruction-following errors, abstract reasoning gaps).

Software Engineering Intern — DevSavant (Bogotá, Colombia)

Jul 2023 – Jan 2024

Built an AI-powered MVP to help elderly users write autobiographies with LLM assistance.
Implemented a Django web app and deployed on a GCP VM for ~10 beta testers.
Integrated the OpenAI API with user-customized prompting; ran work in 2-week sprint tasks.

Selected projects

Using GANs to Detect Cyberthreats — Malware generation + detection with interpretability

Trained a GAN on VirusShare to detect/generate obfuscated and non-obfuscated binary malware samples.
Achieved +3% benchmark accuracy improvement (92% → 95%) vs. a previous CNN approach.
Added interpretability (hooks, GradCAM) and binary reverse-engineering analysis.
Presented at BSIDES Colombia 2024 (cybersecurity conference).

Traffic Signs Recognition Systems — Weather-robust detection with YOLOv8

Fine-tuned YOLOv8 on GTSRB; simulated fog/night/rain via data augmentation.
Modified the final layer and improved mAP50–95: 0.39 → 0.65.
Code

Talks & presentations

BSIDES Colombia 2024: “Using Generative Adversarial Networks (GANs) to Detect Cyberthreats.” (project presentation)

Leadership & service

President — IEEE Computer Society Student Chapter, Universidad del Rosario

Feb 2024 – Dec 2024

Organized internal programming competitions (15–20 participants).
Delivered AI skills workshops for company executives and graduate students.
Led “Code The Future” hackathon (60 students, 15 teams) with UAE Embassy in Colombia; trained teams on Gen-AI fundamentals for sea-level rise solutions; top 5 teams presented to the UAE Ambassador and Dean of Engineering.

Tutor — Universidad del Rosario (Bogotá, Colombia)

Jan 2022 – Jul 2024

Ran 2 weekly sessions (~15–30 students), answered questions, and provided guidance.
Created problem sets and real-world case studies.
Courses: Intro to Data Analysis, Deep Learning (graduate), Differential Calculus, Linear Algebra, Probability & Statistics, Algorithms & Data Structures.

Skills

ML/AI: PyTorch, HuggingFace, Lightning, FastAI · evaluation & failure analysis · interpretability (e.g., GradCAM)
Development: Django, FastAPI, Git/GitHub, Linux, GCP
Math/Tools: MATLAB, mathematical proofs
Languages: Spanish (native), English (C1), Portuguese (B2)

Awards

Fully funded MSc scholarship (UAE government).
50% scholarship (top applicant), Universidad del Rosario.
Recognition of excellence: graduated with top CGPA in cohort.

Contact

Email: jn.sepulveda55@gmail.com
Location/Timezone: Abu Dhabi (Asia/Dubai, UTC+4)