Francesc is currently unavailable
Francesc Guitart
Verified Expert in Engineering
Machine Learning Developer
Lleida, Spain
Toptal member since August 22, 2022
Bio
Francesc is a machine learning engineer and data scientist passionate about building pipelines to automate tasks. He has a PhD in artificial intelligence and 8+ years of experience in the data science sector. He's been working in consultancy firms and technology companies contributing with an innovative attitude and an always-learning spirit.
Portfolio
The Walt Disney Company
Python, PyTorch, Machine Learning, Deep Learning, Data Science...
GFT Technologies
Python, Spark, TensorFlow, Data, Relational Databases...
Eurecat
Python, Scikit-learn, Statistics
Experience
- Artificial Intelligence (AI) - 10 years
- Python - 10 years
- Amazon Web Services (AWS) - 10 years
- Data Science - 9 years
- Machine Learning - 7 years
- Deep Learning - 6 years
- Scikit-learn - 5 years
- PyTorch - 4 years
Preferred Environment
MacOS, PyCharm, Jupyter Notebook, Python, PyTorch
The most amazing...
...thing I've worked on is a suite of machine learning algorithms to understand content in videos.
Work Experience
Senior Machine Learning Engineer | Data Scientist
2019 - PRESENT
The Walt Disney Company
- Developed and integrated machine learning and deep learning pipelines for the extraction of metadata from media content.
- Developed proofs of concept (PoC) and prototypes using computer vision, natural language processing, and multimodal learning for media content automatic tagging.
- Conducted multimodal activity and event detection and classification.
- Developed and integrated an unsupervised algorithm for entity extraction and entity linking of entities from a sports news website, resulting in two patents.
Technologies: Python, PyTorch, Machine Learning, Deep Learning, Data Science, Amazon Web Services (AWS), Data, Relational Databases, Natural Language Processing (NLP), Generative Pre-trained Transformers (GPT), Statistics
Data Scientist
2017 - 2019
GFT Technologies
- Developed a PoC using a deep learning model for the identification and classification of license car plates for a tier 1 Spanish bank.
- Built an NLP model based on Bi-LSTM for the classification of documents and tailored named entity recognition in legal documents for a leading Spanish bank.
- Created a conversational FAQ-based chatbot for a tier 1 Spanish insurance company.
Technologies: Python, Spark, TensorFlow, Data, Relational Databases, Generative Pre-trained Transformers (GPT), Natural Language Processing (NLP), Statistics
Researcher
2015 - 2017
Eurecat
- Developed research initiatives in European and national research funding programs and open calls.
- Collaborated and participated in European and national research projects as an IT and AI partner.
- Contributed to SIM4NEXUS, sustainable integrated management for the NEXUS and Water Innovation Through Dissemination Exploitation of Smart Technologies (WIDEST) projects.
Technologies: Python, Scikit-learn, Statistics
Data Scientist
2014 - 2015
Aplicaciones de Inteligencia Artificial
- Conducted data gathering and cleaning from the National Statistics Institute sources to generate and create a statistical dataset.
- Gathered and cleaned data from real estate portals to create datasets for real estate valuation models.
- Developed real estate valuation models for a leading Spanish bank.
Technologies: R, Statistics
Experience
Multimodal Activity and Event Detection and Classification
Developed a deep learning model for activity and event detection and classification in media content.
• Definition of the problem as multimodal activity, event classification, and timestamp location. The original problem was initially simplified to tackle the top frequent and easiest activity classification within shot boundaries.
• Creation of a multimodal dataset for activity and event detection, classification, and localization. Processed more than five million samples.
• Partnership with PyTorch to develop and integrate a VisualBERT-like multimodal model called MMFTransformer into the MMF project.
• Partnership with PyTorch and Google Cloud to support XLA devices in the MMF project.
• Scale the training of the MMFTransformer model up to a 512 TPU Pod.
• Development of a mixer-based model increased MMFTransformer accuracy by 13%. This enabled automatic activity and event detection, classification, and temporal localization into production according to Disney standards.
BLOG
• https://pytorch.org/blog/how-disney-improved-activity-recognition-with-multimodal -approaches-with-pytorch/
• Definition of the problem as multimodal activity, event classification, and timestamp location. The original problem was initially simplified to tackle the top frequent and easiest activity classification within shot boundaries.
• Creation of a multimodal dataset for activity and event detection, classification, and localization. Processed more than five million samples.
• Partnership with PyTorch to develop and integrate a VisualBERT-like multimodal model called MMFTransformer into the MMF project.
• Partnership with PyTorch and Google Cloud to support XLA devices in the MMF project.
• Scale the training of the MMFTransformer model up to a 512 TPU Pod.
• Development of a mixer-based model increased MMFTransformer accuracy by 13%. This enabled automatic activity and event detection, classification, and temporal localization into production according to Disney standards.
BLOG
• https://pytorch.org/blog/how-disney-improved-activity-recognition-with-multimodal -approaches-with-pytorch/
Multimodal Unsupervised Algorithm for Entity Extraction from Sports News Website
Developed and integrated a multimodal unsupervised algorithm for entity extraction and entity linking of entities from a sports news website.
• Assessment and guidance to define the problem to solve and establish missing or vague requirements, such as evaluation metrics and performance.
• This algorithm can discover and match entities from multimodal unstructured sources of information such as video, audio, and text to a custom taxonomy.
• It enriches article metadata in ESPN content with custom taxonomy entities and its relevance to enable content personalization.
• Processed more than 300,000 articles, handling approximately 600 articles per day.
PATENTS
• https://patents.google.com/patent/US11314706B2
• https://patents.google.com/patent/US11354894B2
• Assessment and guidance to define the problem to solve and establish missing or vague requirements, such as evaluation metrics and performance.
• This algorithm can discover and match entities from multimodal unstructured sources of information such as video, audio, and text to a custom taxonomy.
• It enriches article metadata in ESPN content with custom taxonomy entities and its relevance to enable content personalization.
• Processed more than 300,000 articles, handling approximately 600 articles per day.
PATENTS
• https://patents.google.com/patent/US11314706B2
• https://patents.google.com/patent/US11354894B2
Education
2009 - 2014
PhD in Artificial Intelligence
University of Lleida - Lleida, Spain
2007 - 2009
Master's Degree in Telecommunications
Polytechnic University of Catalonia, BarcelonaTECH - Barcelona, Spain
2004 - 2007
Bachelor's Degree in Computer Science
University of Lleida - Lleida, Spain
Skills
Libraries/APIs
PyTorch, Scikit-learn, TensorFlow
Tools
PyCharm
Languages
Python, R
Platforms
MacOS, Jupyter Notebook, Google Cloud Platform (GCP), Amazon Web Services (AWS)
Storage
Relational Databases
Frameworks
Spark
Other
Artificial Intelligence (AI), Machine Learning, Deep Learning, Data Science, Natural Language Processing (NLP), Generative Pre-trained Transformers (GPT), Data, Statistics, TPU, Custom BERT, Knowledge Graphs, Multimodal Models
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring