Jorge Vanco Sampedro
Jorge Vanco Sampedro

Jorge Vanco Sampedro

I am a Senior Year Student in Mathematical Engineering and Artificial Intelligence. Currently, I am an Exchange Student at the University of Illinois Urbana-Champaign (UIUC).

My research interests focus on Deep Generative Models, Reinforcement Learning, and Large Language Models (LLMs). I am passionate about understanding AI systems from first principles and building efficient, scalable models.

Email / GitHub / Scholar / LinkedIn / X (Twitter)
CV: English · Spanish

Education

Aug 2025 – Present
B.S. Computer Science – Exchange Student (Senior Year) University of Illinois Urbana-Champaign (UIUC)

GPA: 4.0/4
Coursework: Applied Parallel Programming, Deep Generative Models, Reinforcement Learning, Advanced NLP.

Sep 2022 – Present
B.S. Mathematical Engineering and Artificial Intelligence Universidad Pontificia Comillas (ICAI), Madrid

GPA: 9.58/10 · 21 Honors Distinctions.
Coursework: Deep Learning, Autonomous Robots, Big Data, Operating Systems.

Work Experience

Sep 2025 – Present
Research Assistant University of Illinois Urbana-Champaign (UIUC)

Collaborating with the SiNRG Lab (Prof. Romit Roy Choudhury) on Deep Generative Models. Assisted in the research and experiments for a paper on inverse problems and diffusion models, which was accepted at ICLR 2026.

Sep 2024 – Present
Teaching Assistant Universidad Pontificia Comillas (ICAI)

TA for "Foundations of Artificial Intelligence" and "Machine Learning" courses.

Jun 2023 – Jun 2025
Autonomous Vehicle Engineer ISC Racing Team

Involved in the development of an autonomous car to compete in Formula Student competitions.

Publications

Featured Projects

diffusionGPT
diffusionGPT PyTorch MDLM Generative AI
A Masked Discrete Diffusion Language Model (MDLM) chatbot engineered and trained from scratch. This implementation breaks the autoregressive bottleneck by enabling parallel token sampling and leveraging bidirectional context, providing a highly efficient and flexible alternative to standard causal language models.
HELOC Repayment Prediction
HELOC Repayment Prediction Machine Learning XGBoost SVM Explainable AI
A detailed technical report benchmarking credit scoring models for HELOC repayment. Engineered and evaluated diverse architectures, including SVM, XGBoost, and custom Graph-inspired Neural Networks, leveraging SHAP values to ensure model explainability and transparency in financial decision-making.
GPS Madrid System
GPS Madrid System Graph Theory Dijkstra
A custom navigation engine that models the city of Madrid as a complex weighted graph. By applying discrete mathematics and optimized pathfinding algorithms (Dijkstra/A*), this system efficiently calculates optimal routes between urban landmarks, demonstrating practical graph theory application.
LLM from Scratch
LLM from Scratch PyTorch LLMs vLLM
A comprehensive, end-to-end implementation of a Transformer-based Large Language Model. This project covers the entire lifecycle of LLM development, from building a custom tokenizer and designing the attention mechanism to executing distributed pre-training and post-training fine-tuning across multiple GPUs.

Paper Implementations

Stable Diffusion
Stable Diffusion from Scratch
PyTorch U-Net Diffusion
Implementation of the Stable Diffusion architecture (U-Net, VAE, CLIP) built to deconstruct latent diffusion models. Includes a high-performance inference pipeline that optimized v1.5 generation time by over 95% (reducing latency from 8 minutes to 15 seconds).
Generative Models
Generative Models
VAEs GANs Diffusion
An exploration of deep generative modeling focused on pixel art synthesis. Implemented and compared distinct architectures, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Denoising Diffusion Probabilistic Models (DDPMs), to analyze their trade-offs in image fidelity and diversity.
Vision Transformer
Vision Transformer (ViT)
Transformers Computer Vision
Implementation of the Vision Transformer architecture. Implements patch embeddings, multi-head self-attention, and positional encodings to perform image classification without convolutional layers.
Bayesian NN
Bayes by Backprop
Probabilistic DL Uncertainty
Implementation of Bayesian Neural Networks using the Bayes by Backprop algorithm. Introduces weight uncertainty to standard neural networks, enabling the model to express confidence levels in its predictions—critical for robust AI systems.
CNN Explainability
CNN Explainability
CNNs Explainable AI
Implementation of Saliency Maps and Class Model Visualization (Simonyan et al.) to visualize what deep ConvNets actually learn.

Engineering & Applied AI

News Alert Generation
Automatic Alert Generation
NER Sentiment Analysis Multi-task Learning
A multimodal system combining Named Entity Recognition (NER) and Sentiment Analysis (SA) using a joint Bi-GRU architecture. Features a custom uncertainty-based dynamic loss function to balance tasks and integrates LLaMa-3.1 (via Ollama) and ViT-GPT2 for generating context-aware alerts from text and images.
AI Scheduler
AI Scheduler Agent
LLM Agents MCP Google API
An autonomous AI agent built on the Model Context Protocol (MCP) that turns natural language into action. It integrates directly with Google Calendar and Tasks APIs, allowing LLMs to autonomously read schedules, prioritize to-dos, and manage daily planning workflows.
TradeData Architecture
TradeData: Crypto Big Data Pipeline
AWS PySpark Kafka
A comprehensive Lambda architecture for cryptocurrency trading analytics on AWS. Implements a Real-Time layer (TradingView → Kafka → Timestream) for live monitoring and a Batch layer (S3 → AWS Glue/Spark) to calculate complex technical indicators (RSI, MACD) at scale.

View more on GitHub →

Honors & Awards