Mundher Al-Shabi
In my last post, I wrote about using simulators to evaluate multi-turn AI agents. It is a powerful technique, but things can get confusing fast if we do not clearly define the testing architecture. In
Standard vector RAG is great for finding needles, but it’s terrible at describing the haystack. It relies entirely on semantic proximity. That means it often fails at multi-hop reasoning or answering
I've been spending a lot of time lately building autonomous agents with ReAct and RAG frameworks. The shift from stateless chatbots to stateful systems that can actually execute tools is incredibly us
I love the idea of autonomous coding agents. But one of the quickest ways to hit a wall when setting them up is the infinite context problem. Your context window is finite, but the work you want the a
Building LLM apps is easy, but scaling them without setting a pile of money on fire is hard. You really don't need the massive brainpower of GPT-5 for every single user query. Routing is how we fix th
Building a multi-turn conversational AI is surprisingly easy right now. Evaluating it is incredibly hard. For single-turn tasks, a standard static dataset works fine: you just feed in a prompt and ass
I’ve been reading through a recent paper titled "Is Grep All You Need? How Agent Harnesses Reshape Agentic Search". It’s a provocative piece with a premise I normally love. The authors claim that simp
I’ve been spending the last few weeks messing around with open-weight models to build conversational interfaces. By now, the new reality is obvious: generating natural language is no longer the bottle
If you’ve built anything with LLMs in the past couple of years, you’ve probably wired up a Retrieval-Augmented Generation (RAG) pipeline. The playbook is burned into our brains: take a PDF, smash it i