Tag
4 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
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
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
