I like solving challenging problems.

Staffing used to depend on someone manually cross-checking availability, seniority, and past work. The agent does that cross-check on every assignment.

• Built a work-assignment recommendation agent that weighs seniority, fit to the job requirements, scheduled vacations, and semantic similarity to each person's previous work.
• Designed the multi-agent architecture behind the internal chatbot and automated dashboards, so managers query operational capacity and vacations directly instead of assembling reports by hand.
• Automated data analysis for the operational-capacity feature, replacing manual spreadsheet consolidation with dashboards that refresh on their own.

• Deployed local models on virtual machines for production use, keeping inference in-house.
• Used OpenRouter to bring models into features quickly and switch between them without rewriting integrations.
• Worked as a back-end developer across the codebase: code reviews, bug fixes, and maintenance of production services.
• Developed AI features with spec-driven development (Spec Kit / SDD) and Orca to orchestrate coding agents across parallel worktrees.

The first version used Tesseract and coordinate templates annotated in Label Studio: every new provider meant annotating bill areas by hand, and any layout change broke extraction. I rebuilt it around LLMs with mathematical verification.

• Designed a staged pipeline: a cheap model classifies the bill first, then three extractions run in parallel: base data, taxes, and electricity-, water-, or gas-specific data.
• Made steps composable on demand: 5 ready-made pipelines or only the steps a request asks for; if one step fails the others continue and the result comes back as partial.
• Built validation without AI: check digits of the payment line (44-, 47-, and 48-digit formats), barcode amount against the bill total, tax base × rate, item sums, and dates.
• Added automatic correction: when the model gets the payment line or the CNPJ wrong, the system fixes it from the PDF text.
• Hardened it for production: schema-validated output, fallback between providers (Google and OpenAI), and detection of when the model enters a loop.
• Made every call traceable with 7 versioned prompts and cost, tokens, and latency logged per response.
• Exposed it as a synchronous REST API, asynchronous processing through a RabbitMQ queue, and a command-line tool.

Each iteration attacked an observed problem (model loops, request limits, image resolution, digit errors) and was recorded against a comparison answer key covering 3 service types.

• Correctly extracted consumption went from 75% (6/8) to 100% (8/8), and correctly extracted total amount rose by 20 percentage points.
• Failed steps dropped by 100%, average time per bill by 44%, and maximum time per bill from 259.9 s to 45.5 s (−82%).
• Cost per bill fell by at least 54%, and adding a new provider no longer needs any template or manual annotation.
• Ran a feasibility study for open models on own or rented GPUs, estimating US$3 to US$15 per 1,000 bills on a VM with an L40S.

• Built the original Python OCR pipeline with Tesseract on resource-constrained hardware, keeping CPU/RAM usage low.
• Annotated bill fields in Label Studio and trained provider-specific models alongside a generalist model, with automatic recognition of the provider and bill layout.

Research and development fellow on the project “Research on artificial intelligence models applied to legal documents at the Public Ministry of Rio Grande do Norte (MPRN) as support for the development of an Artificial Intelligence Lab project.”

• Built structured extraction for 500+ page legal PDFs with PyTesseract and Azure, cutting extraction time from 15 to 3 minutes per document.
• Parallelized OCR on the GPU, gaining roughly 200% in processing speed.
• Turned the extracted text into input for embeddings that feed an AI pipeline used as a pre-analysis of each case.
• Developed an API for extracting text from legal cases, supporting different extraction methods with Tesseract OCR, PyMuPDF, and FastAPI.

• Worked as a back-end developer using FastAPI.
• Benchmarked Apache Cassandra, MongoDB, and PostgreSQL (vector) on insertion, search, regex search, and vector search in a Docker environment.
• Researched DeepSeek models running on Ollama for use in internal systems.
• Implemented an integration API with Google's AI using FastAPI.
• Studied the Transformer architecture and text representations such as Bag of Words, TF-IDF, and Word2Vec.

• Built the registration and identification module for oncology patients, streamlining hospital access and reducing waiting time at check-in.
• Fixed bugs across the hospital's existing system as a full-stack developer.

• Built public exam pages and notice pages used by candidates to follow each selection process.
• Fixed bugs and maintained the platform in production across front end and back end.