Tag
9 articles
An LLM application fails in ways logs do not capture: the right tool called with the wrong argument, a retrieval that returned nothing useful, a prompt change that quietly broke two other cases. Tracing shows what happened; evaluation tells you whether a change helped.
Rajiv Kumar·Aug 26, 2026·10 min read
LangGraph models an agent as nodes, edges and a shared state object. That structure buys three things a plain agent loop cannot: a place to put state, control over what happens next, and the ability to resume a run that was interrupted.
Rajiv Kumar·Aug 19, 2026·9 min read
An AI coding session ends with a great deal of context in the conversation and none of it anywhere else. A short, structured handoff written at the end is what makes the next session start from where this one finished instead of from the repository.
Rajiv Kumar·Aug 11, 2026·11 min read
An AI assistant opening an unfamiliar repository has to infer everything, and inference produces confident wrong answers. A short brief that the assistant reads before touching anything is the cheapest quality control available.
Rajiv Kumar·Aug 11, 2026·10 min read
A working prompt in a playground is not a system. Production needs a way to tell whether a change helped, a deployment model that matches the traffic, content filtering you have configured deliberately, and observability into what is being spent.
Rajiv Kumar·Aug 5, 2026·10 min read
Copilot Studio makes it easy to build an agent that answers from documents. The gap between that and something people rely on is scope, grounded answers, real actions, and knowing what the agent must refuse to do.
Rajiv Kumar·Jul 8, 2026·11 min read
When a RAG system gives a wrong answer, the instinct is to change the prompt. Almost always the right passage was never retrieved. Chunking, hybrid search, filters and semantic ranking are the levers that actually move quality.
Rajiv Kumar·Jun 22, 2026·11 min read
RAG, evaluation, agents, cost and failure handling are where generative-AI interviews should go. These questions have follow-ups, and the strong answers are about engineering rather than about models.
Rajiv Kumar·May 20, 2026·6 min read
The questions that come up whenever a team starts building with language models: why it makes things up, whether to fine-tune, how much context to send, what it will cost, and how to know whether a change made things better.
Rajiv Kumar·May 6, 2026·6 min read
