AI Engineer | Deep Learning · Computer Vision

NL2SQL Analytics Platform

About the Project

Development of a natural language to SQL (NL2SQL) solution for a Brazilian energy company, enabling business users to query purchase transaction data using plain language instead of writing SQL. The system was built on Vanna AI and deployed on Google Cloud infrastructure within the client's environment. Purchase transaction data was extracted from the client's SAP system and used to train the agent. Users could ask questions in natural language and receive a generated SQL query, the query result, and an automatically produced chart visualizing the data.

• Energy

Responsibilities

- Configured the cloud infrastructure, including deploying the Docker image for Vanna AI within the client's Google Cloud environment. - Trained the agent using purchase transaction data extracted from SAP. Tested and validated query results together with the client's data analysts to ensure accuracy and usability. - Acted as a liaison with Google, reporting on the beta version of Vanna AI and providing feedback on improvements for the tool.

Power Line Insulator Inspection

About the Project

Development of a computer vision solution for an electric energy company to automate the inspection of high-voltage transmission towers using drone-captured images and videos. The system detects different types of electrical insulators, assesses their structural integrity, and, when a defect is identified, classifies the specific type of defect. The solution aimed to replace manual visual inspections with an automated pipeline capable of processing large volumes of aerial imagery, improving inspection speed, consistency, and safety.

• Energy

Responsibilities

- Responsible for planning the overall system architecture (YOLO models), from data ingestion to model deployment. - Trained the computer vision models to detect and classify insulator types and identify structural defects from drone imagery. - Deployed the trained models into production, ensuring the pipeline could process aerial footage and deliver defect classifications reliably.

Reinforcement Learning Game Agent

About the Project

Personal research project exploring reinforcement learning in game environments. Built on top of an custom Jetpack-Joyride-like game clone (built in Pygame), the project involved designing a custom environment and training an autonomous agent to play the game using the Soft Actor-Critic (SAC) algorithm via Stable-Baselines3. The trained model was published on Hugging Face Hub for public access, and the full methodology and results were documented in a technical report.

• GameDev
• AI

Responsibilities

- Designed and configured the reinforcement learning environment based on the game mechanics. - Trained a SAC agent using Stable-Baselines3, tuning hyperparameters and reward structures to optimize agent performance. - Evaluated training results and documented the approach, environment design, and outcomes in a technical report. - Published the trained model publicly on Hugging Face Hub.