Projects and implementations in computer vision, language processing, reinforcement learning, and software development.
LLM finetuning · LoRA · Agents · vLLM
R&D Internship at Orange: building a 20,000+ vulnerability (CVEs) dataset via an agentic pipeline (LangGraph) and LoRA fine-tuning of Qwen 3.6 27B for sovereign SAST code auditing.
RAG Architecture · LangChain · ReAct · Groq API · Pinecone
Design and deployment of an agentic RAG chatbot for this portfolio on a 100% free-tier cloud stack, optimized via targeted 3-chunk per-project breakdown and secured against common attacks.
IEEE ICIP 2025 YOLOv11 · Albumentations · Stable Diffusion
Detection of 14 urban object classes in aerial imagery (IEEE ICIP 2025 / YOLOv11), resolving extreme class imbalance via Data Augmentation (specialized Albumentations and synthetic Inpainting).
Data Engineering · Chained LightGBM · MLP
1st place in Airbus Hackathon: building an AI surrogate model (chained LightGBM & MLP) on 7,500+ features to predict helicopter manufacturing KPIs in real time.
Maskable PPO · Gymnasium
Optimizing dynamic patient-doctor matching via Deep Reinforcement Learning (Maskable PPO), increasing the hospital management score from 45% to 92.5% through semantic action space modeling.
YOLOv8-pose · MediaPipe · Intel RealSense
Development of a 3D computer vision pipeline (YOLOv8-pose, MediaPipe, epipolar triangulation) to reconstruct pen trajectories and capture off-page movements, invisible on graphics tablets, required for dysgraphia diagnosis.
ICPR 2026 Self-Supervised Learning · Vision Transformers · Sentinel-2
Agricultural disease detection in multi-spectral Sentinel-2 satellite imagery: self-supervised pre-training on 125GB of unlabeled data reaching 87.5% accuracy with only 900 labeled samples.
Roadef 2007 Genetic algorithm · Simulated annealing · CP-SAT
Solving an NP-hard telecom technician scheduling problem (ROADEF 2007): benchmarking CP-SAT vs metaheuristics (Genetic Algorithm, Simulated Annealing) across 20 real instances.
PostgreSQL · PostGIS · Python · Folium
Geospatial processing and structuring of 50+ GB of raw GPS telemetry to model real-time traffic for 2,711 buses across 48 routes in Rio de Janeiro.
LightGBM · RandomForest · Time Series
Daily sales forecasting for 54 Ecuadorian supermarkets using time series: feature engineering with oil prices, holidays, and payday cycles, modeled via LightGBM.
LightGBM · XGBoost · RandomizedSearchCV
Credit solvency classification of loan applicants from 42 socio-economic and financial profile variables using optimized gradient boosting models.
LightGBM · Optuna · Basic NLP · Pandas
Real estate price estimation using multivariate regression: amenity feature extraction from free-text descriptions, IQR outlier filtering, and LightGBM model optimization.
XGBoost · GridSearchCV · Feature Engineering
Classification of obesity risk across 7 categories: domain feature engineering and XGBoost optimization via GridSearchCV (90.7% accuracy).
PCA · SVD · K-Means · DBSCAN
Unsupervised segmentation (MovieLens 1M / 6,040 users): outlier purging via K-NN, dimensionality reduction (PCA/SVD), and clustering (K-Means, DBSCAN) revealing distinct film preference personas.
Java · JavaFX · MVC Architecture
Desktop Reverse Polish Notation (RPN) calculator in JavaFX: strict MVC architecture ensuring full plug-and-play view/model decoupling and interoperability.
Java · Swing · XML
Development in Java / Swing of the real-time 2D game Age of War across 3 evolutionary eras: custom rendering engine, dynamic unit stat parsing via XML, and an autonomous AI opponent.
No project in this category.
