Lucas Perez

• Built representation-learning models for similarity search and retrieval of multivariate well-log time series, enabling geologists to find analogous subsurface patterns to support drilling decisions.
• Implemented and tested Transformers, LSTMs, and CNNs to transform raw sensor streams into reusable embeddings for retrieval and downstream analysis.
• Developed a Faster R-CNN-inspired 1D region-p

• Trained and evaluated sparse-autoencoder variants, including Matryoshka and Temporal SAEs, to investigate hierarchical structure in language-model representations.
• Analyzed learned features and model behavior to compare how different SAE architectures recover meaningful hierarchical organization for mechanistic interpretability.

• Oversaw weekly activities for 100+ students across Statistical Foundations of Data Science and Numerical Calculus courses at UFMG.
• Mentored students, developed exercises, graded assignments, answered coursework questions, and supported the preparation of class materials.

• Investigated Fisher Information approximations for continual learning, comparing diagonal and block-diagonal formulations and their effects on model stability and knowledge retention.
• Built and maintained evaluation pipelines to benchmark catastrophic forgetting and retention across continual-learning scenarios.

• Developed an LLM-based pipeline to turn unstructured medical prescriptions into structured fields, combining model outputs with validation and post-processing to reduce manual review.
• Supported an internal OCR system by helping improve data quality and analyzing common failure cases, increasing reliability in production-like inputs.

• Built data-cleaning and preprocessing pipelines for multivariate well-log time series, preparing raw sensor data for similarity-search and retrieval models used by geologists.
• Developed missing-data handling workflows, including imputation modeling using XGBoost, LSTMs and Transformers, to mitigate sensor gaps and ensure consistent inputs for retrieval/embedding training.
• Co-developed a benc

• Graded 500+ second-round OBMEP exams, evaluating mathematical reasoning and written problem-solving.

• Designed and delivered an applied Data Science curriculum covering Python, Pandas, Scikit-learn, and PyTorch.
• Mentored a laboratory-outcome prediction project from problem definition through model evaluation and iteration, supporting faster operational decisions and reduced turnaround time.

• Designed and delivered an applied Data Science curriculum covering Python, Pandas, Scikit-learn, and PyTorch.
• Guided the development of medium-term energy-demand forecasting models to support energy-contracting decisions for industrial operations, with core results published at ABM.

• Built a BeautifulSoup data-collection pipeline covering 1M+ videos and nearly 10K channels from alternative video platforms, including metadata cleaning, validation, and dataset organization.
• Analyzed how YouTube moderation events relate to channel popularity using the collected dataset and YouTube API, contributing to research published at ACM Hypertext 2022.

Co-advisors: Renato Assunção and Fabricio Murai

• Developed Fisher Information-based methods to analyze how pruning reshapes statistical dependencies between neural-network weights, building reproducible PyTorch experiments that led to work published at the ICML 2026 Workshop on Weight-Space Symmetries.

Issued Jul 2024

Issued Mar 2020