Xebia IT Architects
Two months turning RAG prototypes into FastAPI services other people could depend on.
Technical Intern
Xebia is a digital and AI consultancy, which means the question is never whether the model works. It is whether it works on somebody else's Tuesday. I built FastAPI services around LLM APIs so features could be tried in days rather than sprints, Streamlit interfaces so internal teams could query documents conversationally without waiting on a frontend, and LangChain and OpenAI pipelines for the retrieval underneath. Most of what I took away was from the parts that are not the model: chunk sizes, embedding choices, prompt orchestration, and how to tell whether an answer is relevant or merely confident.
What that meant day to day
- FastAPI services wrapping LLM APIs, enabling rapid experimentation with GenAI features.
- Streamlit interfaces for RAG prototypes, used by internal teams against real documents.
- LangChain and OpenAI retrieval pipelines: embeddings, chunking, prompt orchestration.
- Explored evaluation for relevance and hallucination control on enterprise queries.
What I’d point at
I went in thinking the model was the hard part. it was the smallest part. everything difficult was what happens when the model is slow, or wrong, or fine but the request already gave up.