<rafael.macário />
Rafael Macário Fernandes
Open to opportunities

NLP Engineer · Computational Linguist

Rafael Macário Fernandes

I build and evaluate LLM, RAG, and machine-translation systems — with the same rigor for low-resource and Indigenous languages.

São Paulo, Brazil · Portuguese / English / Spanish

01

Selected work

Research 2025

Cross-lingual transfer: Portuguese ↔ Nheengatu

5,028 aligned pairs · fine-tuned XLM-R · P@1 24.7%

Built a 5,028-pair Portuguese–Nheengatu parallel corpus from the Brazilian constitution, then ran four cross-lingual experiments — Word2Vec, fastText, Procrustes/VecMap alignment, and fine-tuned XLM-R — for this Indigenous low-resource language. Fine-tuning lifted cross-lingual sentence retrieval to P@1 24.7% / MRR 0.37, a 22× gain over zero-shot.

Cross-lingual transferXLM-RCorpusLow-resourceEvaluation
First Workshop on NLP for Indigenous Languages of the Americas
Research 2026

USP at AmericasNLP 2026 — culturally-aware captioning

Guaraní — 3rd place in human evaluation (of 8 teams)

Sole-author system (ACL 2026, San Diego) for culturally-aware image captioning across 5 Indigenous languages — Guaraní, Nahuatl, Wixárika, Maya Yucateco, and Bribri. A two-stage pipeline pairs a vision-language model (Qwen3-VL) under cultural prompting with fine-tuned NLLB-200, exposing tokenizer gaps and domain-fit effects.

Qwen3-VLNLLB-200Fine-tuningHugging FaceLow-resource
WiLine AI assistants — public product bot and internal-documentation botWiLine AI assistants — public product bot and internal-documentation bot
Production 2025

Production RAG chatbots for WiLine's products

Production retrieval-augmented chatbots I build end to end for WiLine — both the public in-product assistants on the product sites and the internal RAG bots that answer over the company's internal documentation. Custom retrievers connect a FAISS vector store to LLMs on LangChain for grounded, cited answers, shipped as live chat widgets.

RAGLangChainFAISSLLMsVector searchChatbots
Production 2026

Evaluating RAG systems for hallucinations

WiLine technical white paper: a multi-signal framework that detects hallucinations in RAG chatbots using natural-language inference, semantic similarity, and faithfulness scoring — turning quality signals into a production trust layer.

RAGHallucination detectionNLIFaithfulnessEvaluation
02

Ask my portfolio anything

live RAG chatbot · grounded answers

This is a working retrieval-augmented chatbot — the same kind I ship in production. A transformer embedding model loads in your browser and retrieves the most relevant pieces of my background; a language model then answers your question grounded in exactly those pieces. Expand “sources” under any answer to see what it was built on.

Puranga ára! 👋
Ask me anything 🌿
Rafael's assistant online · grounded in his portfolio
RAG

Puranga ára! 👋 (that's “good day” in Nheengatu) I'm Rafael's assistant 🌿 Ask me anything about his experience, research, or what he can build for you — every answer is grounded in his portfolio, with sources you can check.

03

Experience

Rafael Macário Fernandes

NLP Engineer and PhD candidate in Computational Linguistics at the University of São Paulo. I build and evaluate production-grade chatbots, RAG pipelines, and machine-translation systems — measuring faithfulness, hallucination, and translation quality. My research focuses on cross-lingual transfer between Brazilian Portuguese and low-resource Indigenous languages such as Nheengatu.

NLP Engineer

May 2025 — Present

WiLine Networks · Princeton, US · Remote

  • Built production chatbots and scalable RAG pipelines on LangChain, FAISS, and LLMs — with custom retrievers and APIs into high-availability services.
  • Designed a multi-signal framework to detect RAG hallucinations (NLI, semantic similarity, faithfulness scoring), published as a WiLine technical white paper.
  • Automated multilingual documentation pipelines on GitLab using webhooks, DeepL, and LLMs.
LangChainFAISSRAGEvaluationFastAPIConversational AI

Technical Writer

Sep 2024 — May 2025

WiLine Networks · Princeton, US · Remote

  • Developed and maintained clear, multilingual product documentation with Docusaurus.
  • Extended Docusaurus with custom React/CSS/JavaScript and chat-based assistance.
  • Integrated Google Analytics and Tag Manager to track engagement and optimize content.
React.jsDocusaurusJavaScriptDocumentation UXAnalytics

Linguist Analyst

May 2024 — Sep 2024

Intento, Inc. · San Francisco, US · Remote

  • Ran multilingual machine-translation and AI evaluations across generative tasks — summarization, speech recognition, and more.
  • Built and automated evaluation pipelines, scripting repetitive QA steps in Python.
  • Delivered benchmarking reports and improvement proposals adopted by client-facing teams.
MT/AI evaluationspaCypandassacreBLEUPrompt engineering
05

Documentation, end to end

WiLine Edge Cloud · WEC

WiLine Edge Cloud's documentation, owned end to end — the Docusaurus site (React/CSS), every guide, and the in-product AI assistant. The tutorials below are run from scratch on a real box: the actual commands, versions, and the fixes for what breaks.

Three documentation platforms I design, build, and write — each with its own product identity.

A few highlights — hands-on tutorials & a newsletter, every guide run from scratch on a real box.

06

Skills & tools

Agentic AI & LLMs

RAG & inference

ML & NLP

Data & notebooks

Backend

Frontend

Let's talk

Building something with language & AI?

I'm open to collaborations on NLP, LLM evaluation, RAG systems, and low-resource language technology. The fastest way to reach me is email.

rafael.macario@usp.br