Hallucination detection and response quality at scale.

Shipping anything on top of a language model means answering a question the model cannot answer about itself: is this output actually good? Reading samples does not scale, and a thumbs-up rate tells you almost nothing about why.

This grew out of the LLM-as-a-judge framework I built at Parallel to compare prompt and response quality across models. The dashboard layers hallucination detection through Patronus, topic modelling with LDA, sentiment scoring, and clustering over the embedding space, so quality becomes something you can look at rather than something you argue about.

The point is triage. It shows which prompts drift, which topics produce ungrounded answers, and where quality moves after a change, which is the difference between tuning a system and guessing at it.

What it does

• Hallucination detection via Patronus
• Topic modelling and sentiment across conversations
• Embedding clustering to surface failure clusters
• Built on the LLM-as-a-judge framework from Parallel