Alexander Oh

New York, NY0

AI Research at Petra Labs | Duke AI

AI researcher with a Master's in Artificial Intelligence from Duke, focused on model evaluation, benchmark design, and self-supervised representation learning. Has built retrieval benchmarks, adversarial evaluation suites, and production document parsing pipelines.

AI Fluency

Claude

Coding, learning about deep, complex subjects

Perplexity

Broad search and understanding complex topics

Cursor

Primary Coding Tool (Interactive)

Wisper AI

Efficient Dictation and Speaking practice

Projects

LeDEEP: Depth Perception from JEPA

Ever see a Waymo and wonder why it needs those bulky LiDAR sensors? While modern autonomous systems rely on expensive, specialized hardware, humans navigate the world using monocular RGB input alone. This motivated our development of LeDEEP, a vision-only framework that uses LeJEPA’s self-supervised learning to extract 3D structure from simple camera frames. By leveraging stable SSL objectives to produce rich latent spaces, we’ve built a resource-efficient alternative that shows that robotics don't need a thousand-dollar "hat" to perceive depth.

XGBoost from Scratch

XGBoost is a foundational tool for data scientists, driving some of the most accurate and scalable machine learning solutions across diverse industries. Its tree-based approach delivers efficient, interpretable, and high-performing models, especially suited for structured, tabular datasets. By building XGBoost from scratch, I deepen my understanding of the model and similar algorithms, enabling me to fine-tune performance, interpret results more clearly, and develop future solutions for real-world challenges. This project deepened my understanding of gradient boosting mechanics, optimization strategies, and the inner workings of scalable ML systems.

Advanced Operating Systems Labs

In my operating systems coursework, I gained hands-on experience working directly with kernel-level code to deepen my understanding of core OS concepts. I explored file system internals by navigating inodes, block and fragment addresses, and reimplemented the functionality of Unix commands like find and cat by manually calculating file system structure addresses. Additionally, I experimented with virtual memory management by creating and analyzing custom memory mappings, handling page faults, and observing page swapping behavior using DTrace and C. These projects strengthened my skills in low-level systems programming, memory diagnostics, and performance analysis—key competencies for systems, infrastructure, or performance engineering roles.

ArXivDoc Paper

Document-as-Image Representations Fall Short for Scientific Retrieval https://arxiv.org/abs/2604.18508

Experience

Research Scientist Intern

Petra Labs

Jun 2026 — Sep 2026 · New York, NY

Wrote three research papers on a statistical evaluation framework for Generative Engine Optimization, measuring how prompt variations change LLM citations and the brands named in responses. Advised operators on prompt-map optimization for customer visibility in AI search systems. Built statistical diagnostics that flag significant shifts in citation data for strategy and ops teams.

AI Research Scientist

Duke Natural Language Processing Lab

Sep 2025 — May 2026 · Durham, NC

Showed that semantically guided view selection during ImageNet-1K self-supervised pretraining improves transfer across seven downstream datasets, with the largest gains in low-data settings (+7.6% at 1-shot). Co-led ArXivDoc, an 8,210-document scientific retrieval benchmark comparing document-as-image retrieval with text-based methods. Designed the annotation protocol that narrowed 4,059 candidate queries to 547 evaluation queries. Combined InfoNCE with LeJEPA's SIGReg regularization for self-supervised RL, cutting training time 24% on JaxGCRL Humanoid and AntMaze. Built an adversarial evaluation suite for CLIP under PGD attacks that exposed embedding instability and vision-language failure modes under distribution shift.

AI Capstone Engineer

Huma.ai

Dec 2025 — Jun 2026 · Durham, NC

Architected a LangGraph deep research agent that generates medical affairs documents, saving MSLs 24 to 72 hours of manual research per document. Built a 60-question RAG benchmark over 14 clinical PDFs with a 0.461 baseline that surfaced specific retrieval failures. Created an LLM-as-Judge compliance benchmark grounded in FDA guidance, which showed a 4.5/10 improvement in agent performance. Deployed a table parsing pipeline using UniTable and LayoutLMv3, evaluated with TEDS across more than 10K tables from PubMed research PDFs.

AI Graduate Research Scientist

Duke University

Sep 2025 — Jun 2026 · Durham, NC

Duke Natural Language Processing Research Lab JEPA Models, Reinforcement Learning, Information Retrieval Multimodal AI, Distributed Data Parallel (DDP), Slurm, Adversarial AI (PGD attacks),