Nimesh Bhavsar

Measure before optimizing.

Optimizing against a guess wastes effort. Establish a baseline and a metric before changing anything.

Understand the data before modeling it.

Most modeling failures are data failures in disguise — distribution shift, leakage, or a schema nobody actually read.

Prefer interpretable systems when possible.

A model whose behavior can be explained is easier to trust, debug, and hand off — complexity should be earned, not assumed.

Treat retrieval as an engineering problem, not just an LLM problem.

Chunking strategy, embedding choice, and index design determine RAG quality more than which model sits at the end of the pipeline.

Evaluate AI systems empirically.

"It looks good" is not an evaluation. Define what correct output means for the task, then test against it.