ZenRay is the thing I really wanted when I was debugging RAG and ML pipelines. You run retrieval, reranking, filtering, and then the LLM answers something weird. And you are sitting there like, where did the good candidates get dropped.
So I built ZenRay as an observability layer for ML pipelines. It traces every run and shows what went in and what came out at every stage. It also records why items were dropped, like low score, duplicate, or filtered by rules. That makes debugging feel like a proper science experiment instead of vibes.
It has a Python SDK that you add with a couple of decorators, a FastAPI server that stores traces, and a React dashboard to explore runs and compare them. The goal is zero friction and fast enough to run in production.
Repo: https://github.com/DeepakSilaych/ZenRay