Jonathan Fernandes is a United Kingdom–based independent AI engineer and educator specializing in production-ready retrieval-augmented generation for organizations handling sensitive information. His work translates language-model research into practical retrieval and document-processing systems for banking, insurance, travel, and media.
Fernandes previously worked at language-model companies including Cohere and studied for an MBA at Warwick Business School. By 2022, he was teaching transformers, BERT, and text classification. His subsequent commercial projects included information retrieval for a British bank, publication-specific article summarization, a travel-company retrieval assistant, and document classification and information extraction for an insurer.
His teaching extends to hands-on retrieval-augmented generation and selecting language models, covering embeddings, vector databases, model licensing, privacy, benchmarking, and application-specific evaluation. O’Reilly credits him with teaching more than 250,000 learners.
Building retrieval systems that survive production
- Retrieval quality comes before convincing prose. In a railway-information demonstration, Fernandes shows how an apparently fluent answer can retrieve irrelevant accessibility or travel information. Better embeddings and reranking ultimately direct the user to the correct assistance point.
- Cross-encoder reranking adds accuracy selectively. He uses bi-encoders for fast, scalable first-pass retrieval, then applies computationally heavier cross-encoders to a small candidate set, improving relevance without scoring every document exhaustively.
- Self-hosted RAG infrastructure protects sensitive data. Fernandes prototypes in Google Colab before assembling Docker-based systems with BAAI or NVIDIA embeddings, Qdrant vector search, and locally served Meta or Qwen models through Ollama or Hugging Face Text Generation Inference.
- Observability and retrieval evaluation are deployment requirements. Arize Phoenix or LangSmith exposes latency and pipeline failures; Ragas tests retrieval performance across multiple documents and questions, replacing isolated demonstrations with repeatable assessment.
