RPRajat Pal
← All work

Infosys · for a US bank

GenAI Evaluation Practices

A repeatable way to measure retrieval quality, response quality, and agent behavior, before and after release.

  • LLM evaluation
  • RAG metrics
  • Agent testing
  • Python

The problem

Without evaluation, every prompt change, chunking tweak, or model swap is a guess. Teams couldn’t tell whether a change helped, and regressions reached production unnoticed.

What I set up

Evaluation practices covering three layers:

  • Retrieval quality. Is the right context coming back for a question?
  • Response quality. Is the answer correct, grounded in the retrieved context, and useful?
  • Agent behavior. Does the agent pick the right tools, follow the expected steps, and stop when it should?

How teams use it

Teams run evaluations before release to validate a use case and after release to catch drift and measure improvements. This gave the RAG pipeline tuning work a clear signal for whether a change made things better.