Guim Casadellà
Publications
Guim Casadellà
Data Science, Machine Learning and AI
As a passionate professional in the fields of Data Science, Machine Learning, and AI, I believe in the power of technology to solve complex problems and drive innovation. I’m highly interested in the research, design, and development of advanced solutions.
Below are some of the projects I’ve worked on that best showcase my skills and approach to these fields.
Featured Projects and Experience
1. DatathonFME 2024 WINNING PROJECT - Design Decoder MANGO challenge
- Awards: First prize winners for this datathon’s challenge.
- Overview: Design Decoder is an AI tool aimed at making the life of designers easier by automating the process of logging new garment samples into digital systems. The model extracts design attributes from product images and metadata using a fashion-pretrained CLIP model and traditional AI models for classification.
- Technologies Used: Python, PyTorch, CLIP, XGBoost, OpenCV, Streamlit, NumPy, Pandas, Scikit-learn
- Key Highlights: Image Embeddings, Attribute Classification, Computer Vision, AI Automation
-
GitHub link to the project Kaggle link to Competition
2. Fashion Sales prediction - Machine Learning Course Final Project
- Overview: This Data Science project explores the viability of predicting sales of fashion articles from a multimodal dataset using traditional ML techniques. It develops several techniques to extract insigths from product metadata, images and temporal features.
- Technologies Used: Python, Pandas, PyTorch, OpenCV, Matplotlib, Scikit-learn, XGBoost, SMOTE
- Key Highlights: Image Embeddings, BagOfVisualWords, TimeSeries, Bootstrapping, Bagging, Regression, Classification
-
GitHub link to the project Dataset link
3. Reward Optimization in Semantic Segmentation - CVC Internship
- Institution: Computer Vision Center (CVC)
- Internship Period: January 24 - July 10
- Overview: Developed a reward optimization framework for semantic segmentation tasks during my internship at CVC. This project involved creating algorithms to improve the accuracy and efficiency of semantic segmentation models.
- Technologies Used: Python, Pytorch, TensorBoard, OpenCV, OpenMMLab, DeepLabv3
- Key Highlights: Semantic Segmentation, Reward Optimization, Model Efficiency, Algorithm Development
- GitHub link to the project
4. LoRA for Semantic Segmentation Domain Adaptation - CVC Internship
- Institution: Computer Vision Center (CVC)
- Internship Period: January 24 - July 10
- Overview: Focused on domain adaptation for semantic segmentation using the LoRA (Low-Rank Adaptation) technique. The project aimed to improve model performance when transferring knowledge from synthetic to real-world datasets.
- Technologies Used: Python, Pytorch, SegFormer-B0, LoRA, HuggingFace
- Key Highlights: Domain Adaptation, Semantic Segmentation, Transformer-based Architecture, Low-Rank Adaptation
-
GitHub link to the project HuggingFace link to the models and datasets
5. HackUPC 2024 - Inditex TECH Challenge
- Overview: Developed a solution to detect duplicate or similar images in a dataset, a clothing recommendation system based on the outfit provided by the user, and an assistant system to manage out-of-stock problems for Inditex. Utilized the pre-trained model CLIP for understanding image context and semantic segmentation with a U-NET model for detecting and segmenting clothes in images.
- Technologies Used: Python, Pytorch, Huggingface, CLIP, U-NET, Semantic Segmentation
- Key Highlights: Image Embedding, Contextual Understanding, Clothing Recommendation, Semantic Segmentation, FrontEnd
-
Video Demo GitHub link to the project Devpost
6. CriminalMap
- Overview: Developed a crime mapping app using Python and C. Implemented A* algorithm variation for pathfinding. When computing the best route between two points, it takes in account real time crime data in order to suggest the fastest & safest route.
- Technologies Used: Python, C, Overpass API, Selenium + BeautifulSoup, Google Maps API, Matplotlib
- Key Highlights: Algorithm Design (A*), Data Flow, Data Parsing, Geospatial Analysis, Data Visualization, Web scraping.
- GitHub link to the project
7. NovartisDatathon 2023
- Overview: Developed a ML model for forecasting a company’s sales in the healthcare industry. Focused on Time Series Analysis for extracting information of time-dependent features such as trends, seasons.
- Technologies Used: Python, Pandas, LightGBM, Latitude
- Key Highlights: Time Series Forecasting, Data cleaning, Lag features, Rolling Features, Feature engineering, LightGBM model
- GitHub link to the project
8. Datathon FME 2023 - Outfit generator
- Overview: In the Datathon FME 2023, our team developed an innovative AI tool that recommends outfits aligning with a brand’s core philosophy. The challenge was to integrate diverse data types - from tabular data to image features. Our solution involved merging product data with features extracted using a Convolutional Neural Network (CNN), feeding this rich dataset into a Transformer model to generate brand-aligned outfit suggestions.
- Technologies Used: Python, Pandas, TensorFlow, Matplotlib, Jupyter Notebook
- Key Highlights: Data cleaning, Data preprocessing, Feature extraction, Dimensionality reduction, Fill-in-the-blank, Convolutional Neural Networks, Transformer model, Deep learning
- GitHub link to the project
For more projects, check out my GitHub repositories.
🛠 Skills & Technologies
- Programming Languages: Python, R, C++, C, Dart, Java
- Data Analysis Tools: Pandas, NumPy, SciPy
- Machine Learning Libraries: Scikit-Learn, TensorFlow, Keras, Pytorch
- Deep Learning Frameworks: OpenMMLab, SLM Lab, HuggingFace, OpenAI Gym
- Data Visualization: Matplotlib, Seaborn, Tensorboard
- Database Management: SQL, NoSQL, Firebase
- Software Frameworks: Flutter, Django
- Others: Git, Jupyter Notebooks
🌐 Connect With Me And Check My Other Work
I’m always open to discussing data science and new opportunities. Let’s connect!