Cainã – Python, LLM, LangGraph

A US-based online education streaming platform that offers high-production video courses taught by famous experts, celebrities, and industry leaders. The company launched a new product called "Certificates", where they partner with renowned companies like Microsoft, Google, and NVIDIA to create high-quality training certificates. Each certificate has one or more capstones that should be graded using AI. A workflow was developed to automate the AI Evaluator validation and optimization.

• Developed and automated the whole pipeline to inspect, validate, and optimize the existing AI Evaluator of the certificates' programs;
• Developed and deployed UI apps on HuggingFace Spaces that help learning designers to validate the course rubrics and optimize the assessment questions;
• Structured the project as a Python subpackage from core functions to CLI commands.

A text-to-SQL AI Assistant integrated on Slack, with human-in-the-loop before query execution. The AI Assistant can either generate and execute SQL code (under user approval) or use specific tools. The toolkit includes the ability to log Q&A examples used on RAG and improve the code generation accuracy (for data engineers) or a careful didactic explanation of the query (for stakeholders who want to understand SQL or the codebase).

• Implemented the end-to-end text-to-SQL AI Agent from scratch, including Python core functions, FastAPI endpoint with gateway, Slack webhook;
• Validated LLM outputs with SME feedback analysis;
• Adopted the best Python practices using SOTA approaches.

An AI-powered cattle mapping system to combat Amazon deforestation. The system downloads and scans high-resolution satellite images (e.g., from protected or embargoed areas) and detects and counts cows.

• Led a team of data science undergrad students;
• Downloaded and annotated >10k images and >100k cattle;
• Trained and optimized deep learning models to estimate cattle counts;
• Built a scalable system that can be applied to target regions like protected and embargoed areas.

An open-source Python package developed as part of the postdoc at UW-Madison that extracts and maps SQL dependencies and outputs them using Large Language Models (LLMs). It is 150X faster and 400X cheaper than human manual inspection. It scans for the SQL codebase and returns a structured output with:

1. Dependencies: Tables and columns that must exist BEFORE query execution
2. Outputs: Tables and columns permanently CREATED or MODIFIED by the query while ignoring temporary constructs. The key benefits include safe database schema updates, database storage optimization, database migration plans, visual schema understanding, and comprehensive database documentation.

• Implemented the end-to-end Python Package from scratch;
• Developed the core Python API as well as CLI and Web interfaces;
• Validated LLM outputs using a ground truth dataset;
• Adopted the best Python practices using SOTA approaches (uv, pytest, CI/CD, versioning, containerization, etc).

Digital slickline (DSL) has been introduced to improve the efficiency of intervention operations in both onshore and offshore wells. Most nonproductive time (NPT) in DSL operations stems from telemetry issues caused by cable faults, despite the system's robustness. To address this, a data-driven framework was developed to identify potential cable damage and its approximate location using DSL logging telemetry data. The method was tested on 992 real-world downhole jobs across almost 30 countries. To validate, method predictions were compared for 60 jobs with labeled potential faults (i.e., cable damage), achieving 98% accuracy in determining whether each job had a fault.

• Deployed multiple machine learning models to predict asset failures and optimize maintenance schedules;
• Designed explainable risk indices using regression and clustering;
• Published a statistically based model for detecting failures.