Thoughts on building intelligent systems, AI engineering, and the craft of making machines understand context.
Most RAG tutorials stop at 'chunk your docs and query.' Real production systems need hybrid retrieval, re-ranking, semantic caching, and graceful fallbacks. Here's the architecture I use to build RAG pipelines that actually work at scale.
Read articlePrompt engineering is dead. The real leverage is in what you feed the model before the prompt—system instructions, retrieved documents, conversation history, tool outputs. Here's how to think about context as a first-class engineering problem.
Read articleWhen I trained ModernPubMedBERT, generic embedding models failed on medical terminology. This post covers the full pipeline: curating contrastive pairs, choosing loss functions, avoiding catastrophic forgetting, and benchmarking.
Read article