AI Deployment: RAG, Fine-Tuning and Hugging Face

Deploy an AI system that other people rely on: measure what a request costs in tokens, cedis and mobile data; build hybrid retrieval with reranking and citations rather than a single vector search; prove groundedness with an evaluation that gates a merge; adapt a small model when the missing thing is behaviour rather than knowledge; and serve, quantise and price it on hardware you can actually get. Ends with a source-linked NHIS medicine-information assistant, bounded to informational answers and defended in writing. Assumes C1 and comfort with Python. Reviewed: 2026-09-10