Real-World Headaches in Production LLM Systems: From Manual to Automated Evals

JuniorLeg6988 · reddit · 2026-08-12

A developer initiated a discussion on the specific, boring failures encountered when building production-grade LLM systems and RAG pipelines, noting that failure handling often relies on duct-tape fixes.

The author's case study: The team used to evaluate agentic systems manually, taking 5-10 hours per run. To automate this, they used browser automation to drive the agent, fetched ground truth from Snowflake, and passed the context to Claude 3 Opus to generate pass/fail evaluations.

Key takeaway: The AI judge wasn't the final word. It generated a report for human review, cutting review time to 30-60 minutes by removing mechanical tasks. The ongoing challenge remains how to responsibly close the final gap of auditing the AI judge without simply defaulting to another layer of human review.

Original post →

More from coding & agent

coding & agent channel →