Rafael is currently unavailable
Rafael Toledo
Verified Expert in Engineering
Computer Vision Developer
Florianópolis - State of Santa Catarina, Brazil
Toptal member since April 24, 2019
Bio
Rafael has been working in artificial intelligence for almost 15 years. Half of this time has been focused on computer vision, where he developed solutions for background removal apps, image matting, OCR, color detection, face manipulation, and autonomous vehicle perception. His work involves developing real-world applications from scratch. In production, Rafael has built solutions with PyTorch, TensorFlow, TensorFlow Serving, and Flask, using cloud resources from Azure, AWS, and Google Cloud.
Portfolio
Pixelcut Inc.
PyTorch, NVIDIA TensorRT, Tritonserver, Python, Semantic Segmentation...
MIME
Python, TensorFlow, PyTorch, Git, Docker, Amazon Web Services (AWS)...
Athena Analytics
ARIMA, XGBoost, PSQL, Python, Data Science, Machine Learning, Flask, Docker...
Experience
- Deep Learning - 15 years
- Computer Vision - 15 years
- Machine Learning - 15 years
- Python - 8 years
- OpenCV - 7 years
- Image Processing - 7 years
- TensorFlow - 5 years
- PyTorch - 5 years
Preferred Environment
Python, TensorFlow, PyTorch, Amazon Web Services (AWS), Computer Vision, Deep Learning, OpenCV, Git, Azure, Flask
The most amazing...
...project I've worked on was an autoML that trained, built, and deployed optimum models remotely and handled cloud resources on demand, achieving 98% accuracy.
Work Experience
Computer Vision Engineer
2022 - PRESENT
Pixelcut Inc.
- Developed multiple BGRM architectures models with a robust performance addressing types of scenes, such as jewelry, mannequins, garments, logos, people, and matting properties for hair and fur strands.
- Deployed deep learning models using TensorRT and Triton server.
- Built solid, massive, diverse, and high-quality datasets to address the BGRM model needs.
Technologies: PyTorch, NVIDIA TensorRT, Tritonserver, Python, Semantic Segmentation, Image Matting
Computer Vision Engineer
2019 - 2022
MIME
- Built and trained a custom neural network to predict skin tones and complexions from selfies, achieving more than 80% accuracy, a very high standard for color. It followed a server deployment that answered more than 2,000 predictions daily.
- Adapted a computer vision algorithm to estimate a scene illumination, using spherical harmonics radiance mapping and face 3D mesh, relighting the faces afterward.
- Built a face skin segmentation model that distinguished skin from face members for foundations virtual try-on. It was gradient-boosting based and achieved more than a 95% accuracy.
- Adopted cutting-edge deep learning algorithms like Variational AutoEncoder (VAE), Weak Supervised Learning, Physics-guided learning, and built custom architecture with multiple custom losses and outputs.
- Deployed multiple applications for production on AWS, using EC2, S3, Elastic Beanstalk, Elastic Inference, and TensorFlow Serving.
Technologies: Python, TensorFlow, PyTorch, Git, Docker, Amazon Web Services (AWS), Data Science, Machine Learning, Deep Learning, Computer Vision, Flask, OpenCV, Artificial Intelligence (AI), Computer Vision Algorithms, AWS Elastic Beanstalk, Amazon EC2, XGBoost, Amazon S3 (AWS S3), Gradient Boosting, Gradient Boosted Trees, Supervised Learning, Labeling, Bash, SSH, Jupyter Notebook, Jupyter, Variational Autoencoders (VAEs), R, Linux, REST APIs, Pandas, NumPy, RStudio, Functional Programming, Tidyverse, Vim Text Editor, Tmux, Decision Tree Regression, PSQL, Image Processing, Convolutional Neural Networks (CNNs), Data Analysis, Data Visualization
Data Scientist
2019 - 2020
Athena Analytics
- Invented multiple new grade features and developed an autoregressive model that predicted grades from more than 100,000 students with a 93% accuracy in many Irish schools.
- Created an autoML system that generated optimum new models as more samples were added. The optimization was based on the hyper-tuning bayesian search.
- Automated crontab service to feed predictions on the database as more data were added.
Technologies: ARIMA, XGBoost, PSQL, Python, Data Science, Machine Learning, Flask, Docker, Bash, SSH, Jupyter Notebook, Linux, Git, REST APIs, Pandas, NumPy, RStudio, Functional Programming, Tidyverse, Vim Text Editor, Tmux, Decision Tree Regression, Gradient Boosted Trees, Artificial Intelligence (AI), Gradient Boosting, Supervised Learning, Jupyter, Data Analysis, Data Visualization
Data Scientist
2016 - 2019
Khomp
- Developed an AI service that distinguished human voices from machine/records voices with 98% accuracy. This service predicted more than 150,000 calls daily for tens of customers on production.
- Developed an autoML service for independent custom training of clients that delivered optimum models in less than one hour, requiring them only to send data to the cloud.
- Implemented an autoML service that instantiated resources only by demand on Azure. Everything (Blobs and Batch Computing) on automatically instantiated and automatically destroyed.
- Implemented a biometric voice system for more than 100 people from scratch using Gaussian mixture models, gradient boosted tree models, hypothesis tests, and a custom file database system.
- Trained models to recognize a speaker's emotion, age, and gender, using the Gaussian mixture and gradient boosted trees models.
- Implemented Continuous Integration and Continuous Deployment (CI/CD) configurations in five GitLab projects and the webhooks in the Docker registries.
Technologies: Continuous Delivery (CD), Continuous Integration (CI), Azure, Git, Docker, Python, R, Data Science, Machine Learning, Flask, Azure Blobs, Blob Storage, XGBoost, Gradient Boosting, Biometrics, Emotion Recognition, Speech Recognition, Gradient Boosted Trees, Jupyter, Jupyter Notebook, Bash, SSH, Linux, RStudio Shiny, REST APIs, Pandas, NumPy, RStudio, Functional Programming, Tidyverse, GitLab CI/CD, Digital Signal Processing, Vim Text Editor, Tmux, Decision Tree Regression, PSQL, Artificial Intelligence (AI), Supervised Learning, Labeling, Data Analysis, Data Visualization
Computer Vision Engineer | Data Scientist
2014 - 2016
Neoway
- Trained more than 200 CAPTCHA's recognition solutions, achieving performance superior to 80%. Used classical approaches with SVM and a deep learning approach with VGG19 and Caffe.
- Applied reverse engineering to formulate numbers from more than 50 Brazilian government registers of open sources to internet bots.
- Formulated hundreds of specific queries on PostgreSQL databases to answer customers' questions and analyses automatically.
- Conducted tens of statistical analyses of internal services to detect bottlenecks and failures and to guide the subsequent engineering decisions.
Technologies: Java, PostgreSQL, Git, Docker, Python, R, Data Science, Machine Learning, Deep Learning, Computer Vision, Flask, SVMs, Computer Vision Algorithms, Amazon Web Services (AWS), SSH, Bash, Image Processing, Caffe, OpenCV, Linux, RStudio Shiny, REST APIs, Pandas, NumPy, RStudio, Functional Programming, Jupyter Notebook, Tmux, PSQL, Artificial Intelligence (AI), Convolutional Neural Networks (CNNs), Amazon EC2, Amazon S3 (AWS S3), Supervised Learning, Labeling, Data Analysis, Data Visualization
Experience
Face Reconstruction with Variational Autoencoder and Face Masks
https://github.com/tldrafael/FaceReconstructionWithVAEAndFaceMasksThis project accomplishes a paper publication with the same name on ENIAC 2021 (National Conference of Artificial and Computational Intelligence).
This work investigated how face masks can help the training of VAEs (variational autoencoders) for face reconstruction by restricting the learning to the pixels selected by the face mask. An evaluation of the proposal using the CelebA dataset shows that the reconstructed images are enhanced with the face masks, especially when SSIM loss is used either with L1 or L2 loss functions.
This work investigated how face masks can help the training of VAEs (variational autoencoders) for face reconstruction by restricting the learning to the pixels selected by the face mask. An evaluation of the proposal using the CelebA dataset shows that the reconstructed images are enhanced with the face masks, especially when SSIM loss is used either with L1 or L2 loss functions.
Facial Attributes Manipulation with Variational Autoencoder
https://github.com/tldrafael/ManipulateFacialAttributesUsingVAEIt was modeled and trained a Variational Autoencoder (VAE) that emulated different facial attributes on subjects' photos like hair color, age, beard, mustache, smiling, and glasses-wearing.
The project enables to augment a small face database and make the same face available in many distinct expressions. This augmented data can help in the training of other algorithms.
It is based on the CelebA dataset that contains more than 40 facial attributes with labels.
The project enables to augment a small face database and make the same face available in many distinct expressions. This augmented data can help in the training of other algorithms.
It is based on the CelebA dataset that contains more than 40 facial attributes with labels.
Dashboard for Urban, Rural, and Indigenous Populations in Brazil
https://github.com/tldrafael/brazil_urban_rural_indigenous_censusThis project explored Brazil's census data with an R and Shiny library. It presents a dynamic dashboard making it possible to see the evolution of these three populations over the years and where precisely these populations are across the national territory.
The application can be found at the link below:
• https://toledo-rafael.shinyapps.io/urban-rural-indigenous-census-BR/
The application can be found at the link below:
• https://toledo-rafael.shinyapps.io/urban-rural-indigenous-census-BR/
Azure Services Presentation
https://github.com/tldrafael/AzurePresentationWithShinyAppIn the presentation, I explained how to use services from Azure. An application was created to train a model and make predictions from it. The model predicts who composed the song, the Beatles or the Rolling Stones.
The presentation is presented at https://github.com/tldrafael/AzurePresentationWithShinyApp/blob/master/helpers/AzurePresentation.pdf
The presentation is presented at https://github.com/tldrafael/AzurePresentationWithShinyApp/blob/master/helpers/AzurePresentation.pdf
Implementing a Neural Net from Scratch
https://github.com/tldrafael/implementing-MLP-from-scratchIn this project, I have implemented a simple Multilayer Perceptron (MLP) from scratch for educational purposes.
The project includes gradient computation, gradient-checking, and the backpropagation algorithm.
The project includes gradient computation, gradient-checking, and the backpropagation algorithm.
Education
2023 - 2026
Doctorate Degree in Computer Vision
Federal University of Santa Catarina - Florianópolis, Brazil
2020 - 2022
Master's Degree in Computer Vision
Santa Catarina Federal University - Florianópolis, Brazil
2007 - 2011
Bachelor's Degree in Mechatronics Engineering
Amazon State University - Amazonas, Brazil
Certifications
OCTOBER 2021 - PRESENT
Generative Adversarial Networks (GANs) Specialization
Coursera
MAY 2020 - PRESENT
TensorFlow: Data and Deployment Specialization
Coursera
APRIL 2020 - PRESENT
TensorFlow in Practice Specialization
Coursera
JULY 2019 - PRESENT
Deep Learning Specialization
Deeplearning.ai via Coursera
OCTOBER 2016 - PRESENT
Developing Data Products
Coursera
SEPTEMBER 2016 - PRESENT
Practical Machine Learning
Coursera
JUNE 2016 - PRESENT
Machine Learning
Coursera
JANUARY 2016 - PRESENT
Regression Models
Coursera
OCTOBER 2015 - PRESENT
Statistical Inference
Coursera
MARCH 2015 - PRESENT
Image and Video Processing: From Mars to Hollywood With a Stop at the Hospital
Coursera
FEBRUARY 2015 - PRESENT
Reproducible Research
Coursera
DECEMBER 2014 - PRESENT
Getting and Cleaning Data
Coursera Course Certificates
DECEMBER 2014 - PRESENT
Exploratory Data Analysis
Coursera
NOVEMBER 2014 - PRESENT
The Data Scientist's Toolbox
Coursera
NOVEMBER 2014 - PRESENT
R Programming
Coursera
Skills
Libraries/APIs
XGBoost, TensorFlow, PyTorch, OpenCV, Tidyverse, Pandas, NumPy, Keras, REST APIs
Tools
Git, Jupyter, ARIMA, Vim Text Editor, GitLab CI/CD, Tmux
Languages
Python, R, Bash, Java
Frameworks
Flask, RStudio Shiny, Caffe
Platforms
Jupyter Notebook, Amazon Web Services (AWS), Docker, RStudio, Debian, Linux, AWS Elastic Beanstalk, Amazon EC2, Azure
Paradigms
Continuous Integration (CI), Functional Programming, Continuous Delivery (CD)
Storage
PostgreSQL, Amazon S3 (AWS S3), Azure Blobs, PSQL
Other
Gradient Boosted Trees, Artificial Intelligence (AI), Machine Learning, Data Science, Deep Learning, Computer Vision, Computer Vision Algorithms, Image Processing, Convolutional Neural Networks (CNNs), Variational Autoencoders (VAEs), Data Analysis, Data Visualization, Speech Recognition, Blob Storage, Gradient Boosting, Biometrics, Emotion Recognition, Supervised Learning, Labeling, Decision Tree Regression, Engineering, SSH, Digital Signal Processing, Natural Language Processing (NLP), Generative Adversarial Networks (GANs), SVMs, Neural Networks, Generative Pre-trained Transformers (GPT), NVIDIA TensorRT, Tritonserver, Semantic Segmentation, Image Matting
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