I build practical AI systems that connect language models with real products, messy data, and business workflows.

Website · Email · GitHub

I'm a Brazilian AI Engineer working mainly with LLMs, RAG systems, agentic workflows, evaluation, and applied NLP.

My background is a mix of production AI engineering and NLP research. I have worked on GenAI systems for banking, fintech, document automation, retrieval, and multilingual language model evaluation.

I like building AI systems that are not just demos: systems that need to retrieve the right data, behave reliably, reduce cost, improve workflows, and survive contact with real users.

Some areas I especially enjoy:

• Building RAG systems over complex or messy business data
• Designing evaluation loops for LLM and retrieval systems
• Creating agentic workflows with LangChain / LangGraph-style architectures
• Turning research ideas into usable products
• Working across product, engineering, and business constraints

💼 Built GenAI and RAG systems for companies in banking, fintech, and enterprise document automation.

📉 Helped reduce AI system costs and improve retrieval/evaluation quality in production-oriented environments.

📚 Creator of Napolab, a Portuguese NLP / LLM evaluation benchmark and dataset collection.

🔎 Creator of hashformers, a library for multilingual hashtag and word segmentation recognized at LREC 2022.

🤗 Former contributor to Argilla, later acquired by Hugging Face.

Portuguese language model evaluation benchmark and dataset collection.

Napolab was created to study how Portuguese language models behave across different tasks, datasets, and evaluation settings. One of its key contributions is FaQuAD-NLI, which has been reused by the Portuguese NLP community in evaluation tooling and leaderboards.

• Napolab Leaderboard
• Master’s thesis: Lessons learned from the evaluation of Portuguese language models
• Article: The Hidden Truth About LLM Performance

Research code for multilingual hashtag and word segmentation.

Hashformers uses language models and beam search to segment hashtags and whitespace-free text. The project was recognized as state-of-the-art for hashtag segmentation at LREC 2022 and has been cited or reused in later research.

• Paper: Zero-shot hashtag segmentation for multilingual sentiment analysis

Legal NLP research code for retrieval and entailment.

• To Tune or Not To Tune? Zero-shot Models for Legal Case Entailment
• Yes, BM25 is a Strong Baseline for Legal Case Retrieval

During my internship at Argilla, I worked on open-source NLP tooling related to annotation workflows, weak supervision, and embedding-based features.

Argilla was later acquired by Hugging Face.

Selected contributions to ML, NLP, and research tooling:

• Hugging Face Transformers: modified the Trainer class to support simultaneous Ray Tune and Weights & Biases execution.
• Facebook Research BLINK: fixed a parameter bug in a benchmark script.
• Portuguese Word Embeddings: fixed an evaluation bug later documented in a published paper.
• AWS Labs mlm-scoring: improved installation instructions for the library.

I occasionally write about LLM evaluation, NLP benchmarks, and language model behavior.

• The Hidden Truth About LLM Performance
• Master’s thesis: Lessons learned from the evaluation of Portuguese language models
• Google Scholar

I'm especially interested in roles involving:

• AI Engineering
• LLM applications
• RAG systems
• Agentic AI
• LLM evaluation

You can reach me here:

• Website: ruanchaves.github.io
• Email: ruanchaves93@gmail.com