Thiago – PyTorch, Python, Machine learning

A platform that connects talented individuals with the right job opportunities. For this project, Thiago designed an intelligent matching algorithm that automatically categorizes talent-to-job matches as bad, okay, good, or excellent. With this system in operation, the matching process becomes more efficient. Instead of manually reviewing each application, we can suggest suitable candidates for jobs and vice versa.

• Made extensive research for the best approach to the problem and documented it;
• Developed the smart-matching algorithm;
• Achieved a 300% gain in the main metrics;
• Developed production code with ETLs for data and Feature Store, deployed model on AWS Sagemaker.

A pioneering venture aimed at leveraging AI to enhance Quality Assurance. Its flagship product, FastQA, equips businesses with the tools to effortlessly automate tests for their APIs and Web Applications. Thiago played a pivotal role in developing the backend and the AI models, and also took on roles as a product and engineering manager. Currently, he is not actively engaged with RedwolfAI on a full-time basis, as his continuous presence is not required.

• Created an AI system that makes it easy to create automated Web App tests;
• Created the Django backend for the whole application;
• Created Azure Pipelines to ensure continuous integration, continuous deployment, and continuous testing;
• Deployed the system on Microsoft Azure;
• Participated in pre-sales.

A Los Angeles-based company specializing in crafting AI modules for their web application. Thiago developed an AI solution that skillfully extracts and standardizes data from tables in PDFs, regardless of their diverse layouts and formats. He trained a model to recognize table structures and seamlessly integrated ChatGPT to transform the extracted data into a uniform format.

Finally, the AI system was deployed using FastAPI and Docker.

• Successfully implemented the AI pipeline - it works with several different types of tables from various PDF files;
• Wrapped the AI pipeline in a FastAPI API so that other systems can communicate with the AI;
• Dockerized and deployed the system on AWS.

One of the largest engineering companies in Brazil, responsible for numerous highways across the country. In this projec, Thiago developed two systems: Traffic Counting and LVC.

• Traffic Counting: Thiago spearheaded the development of a robust Machine-Learning framework capable of counting and categorizing vehicles under diverse conditions, encompassing various states, weather patterns, and time settings. This model exhibited exceptional performance in analyzing video footage, whether it was under the bright sunshine or amidst a rainy night. Moreover, the dev team met the client's critical requirement for real-time processing, achieving seamless, instantaneous results. In addition to model development, Thiago oversaw the entire data collection and annotation process, a crucial step in accurately categorizing vehicles. For instance, when dealing with trucks, they not only identified the truck itself but also determined the number of axles it possessed, delivering comprehensive and precise results for the client.

• LVC: an AI system that counts and categorizes flaws in highways. This cutting-edge system not only accurately enumerates and classifies these flaws but also provides precise information regarding their specific location within the lane (left, center, or right) and estimates the flaw's length in meters.

• Automated the vehicles and flaws counting/categorizing process, making it possible for Strata to process many more highways, as the code can run without breaks and never gets tired, with a fraction of the cost;
• Implemented a simple GUI to make it easier to run the code.

Thiago spearheaded a project aimed at identifying users with a high propensity to purchase a vehicle. The objective was to seamlessly direct these high-intent users to their affiliated financial institutions for car loan options. This initiative generated revenue for the company through a partnership model, whereby the company earned a commission each time a user engaged in a car financing simulation via the platform.

• Developed a sophisticated predictive model to identify users most likely to purchase a vehicle accurately;
• Implemented a system that automatically triggered targeted emails and push notifications for high-intent users;
• Designed and maintained a specialized interface for simulating car loan options;
• Significantly improved the engagement rate for notifications from approximately 2% to an impressive 11%.

A predictive analytics model was designed to pinpoint clients at an elevated risk of churn for a financial institution. The primary objective of this endeavor was to enable proactive retention strategies. Once the model flagged clients as high-risk for attrition, the bank could then initiate targeted interventions to enhance customer loyalty and reduce churn.

• Developed and implemented a predictive analytics model;
• Established quantitative metrics for tracking user retention;
• Demonstrated the model's performance improvement, resulting in the prevention of approximately $2 million in asset loss;
• Received recognition from business stakeholders for insightful predictions and contributions to client retention efforts.

In this project, Thiago created a deep-learning model capable of identifying skin cancer from dermatoscopic images. The idea was to train a CNN to large amounts of high-quality data in order to have a trained model that generalizes to new images.

They crafted a data collection/data labeling pipeline to include data from Brazil specifically, as the dataset we were using was mainly for caucasian people (Brazil has a lot of brown people).

After training the CNN, they wrapped it into a FastAPI API so that the dev team could use the CNN on the Android app that was created to help the physicians in the project.

• 80% F1-score on the validation data
• Model deployed and successfully made available as an API for the dev team.

Thiago designed an advanced algorithmic system aimed at identifying and categorizing similar pieces of legislation based on a given sample. This initiative was part of the Agency's broader campaign to streamline its existing legislative corpus. The first critical step toward this objective was to cluster related legislative texts, enabling the consolidation of these texts into unified, simplified documents for easier navigation and interpretation

• Leveraged Deep Learning and Natural Language Processing to enhance our search engine's capabilities. Instead of merely fetching documents based on keyword similarity, the engine now intelligently identifies and retrieves content that aligns with the subject matter of the query;
• Developed a comprehensive PowerBI dashboard, facilitating agency-wide access and utilization of the newly advanced search engine, thereby democratizing data-driven insights and decision-making across the organization.

In this project, the goal was to segment satellite images in order to find central pivots. Central pivot is an irrigation technique that is crucial for agriculture, especially when it comes to large-scale production. By identifying central pivots, the team could estimate crop productivity and water use, find illegal planting areas, etc. In this project, the team used Pytorch to train a U-NET model using the 7-channel satellite images.

• 97% Average Precision
• Inference code deployed in an internal server so that people from the lab could use the model effortlessly.