I'm Emiliano Romero, a Computer Engineering student and AI Researcher. This space is an open repository for experiments, technical writeups, unfiltered learnings, and ideas that don't fit into a standard CV.
Self-supervised visual encoders, feature distillation, and geometric priors in foundation models.
Low-latency inference engines, CUDA acceleration, and deploying deep learning on constrained edge hardware.
Intensive summer coursework in advanced Computer Vision at Oxford University, covering spatial and generative models.
Unfiltered engineering notes, papers I'm reading, mental models, and learnings outside LinkedIn's character limit.
Investigating methods to transfer high-dimensional spatial representations from large vision foundation models (DINOv2) into sub-50ms convolutional backbones for resource-constrained robotic platforms.
Benchmarking latency and memory bottlenecks across multi-head attention kernels in visual transformers. Evaluated custom memory layouts to maximize tensor core saturation under variable image patch resolutions.
An exploration into contrastive visual-language embeddings to flag out-of-distribution physical obstacles without supervised defect datasets. Designed for fast inference on ARM architectures.
A minimalist distributed tracking tool built for personal model training runs. Provides live telemetry over WebSockets with zero external cloud SaaS dependencies and micro-second logging overhead.
“Studying Computer Vision at Oxford reshaped how I think about perceptual representation in neural architectures. Moving beyond black-box classification towards 3D spatial reconstruction, multi-view geometry, and generative priors.”
Most modern video architectures treat time as an extended spatial dimension rather than a directional causal flow. Notes on why autoregressive tokens fail to preserve physical object permanence across occlusions.
A collection of personal takeaways from Oxford lectures: how Euclidean distance metrics fall short in semantic high-dimensional manifolds, and why spherical embeddings matter for contrastive loss functions.
Premature optimization is a distraction, but knowing your hardware memory hierarchy is non-negotiable. Reflections on profiling GPU memory transfer bottlenecks before touching any kernel code.
Why domain randomization often beats hyper-realistic rendering pipelines when training agents to navigate dynamic, noisy real-world environments.
I am an undergraduate student in Computer Engineering with a rigorous focus on Artificial Intelligence and Computer Vision. My work sits at the intersection of statistical machine learning and high-performance systems engineering.
During my summer programme at the University of Oxford, I immersed myself in the mathematical and perceptual foundations of computer vision. I am interested in building vision systems that understand not just pixels, but geometry, physics, and causal mechanics.
This website serves as my public sandbox: a minimalist space to deploy prototypes, document research iterations, publish reading lists, and share things that standard professional profiles cannot accommodate.