Have you ever had that...
Have you ever had that feeling – after pushing your model to production, your heart sinks even more?
I totally get it. Before launch, the metrics on the test set looked beautiful, and your boss nodded in approval. But once it's deployed to the real world and users start using it, who knows what could happen. The data distribution quietly shifts, input formats go haywire, and the model starts spitting out some utterly baffling responses – and you have absolutely no idea where things went wrong.
What's worse, you don't even know that it doesn't know. LLMs aren't like traditional programs that crash with an error – they just confidently spout nonsense. You ask it about company policy, it fabricates an answer with total conviction. The user believes it, and then the complaints come rolling in.
This is the "non‑deter...
This is the "non‑deterministic failure" of AI systems – the ways they break are things traditional software would never do.
When I first got into MLOps about two years ago, I was on the lookout for tools that could address this issue. Then I came across Evidently AI, an open‑source framework for AI evaluation and observability. To be honest, I didn't have high expectations at first – I'd tried plenty of open‑source monitoring tools, and they were either too weak in functionality or so complex to configure that they made me want to throw my keyboard across the room.
But Evidently is different.
Its core logic is actu...
Its core logic is actually quite simple: you can't trust what you haven't tested. So it provides a full toolchain that lets you run quality checks and compare prompts and model performance during development; run adversarial stress tests before deployment; and continuously monitor output quality once you're in production.
It's not a hassle to use either. It's just a Python library – you give it your data (pandas DataFrame or Spark works) and it generates visual evaluation reports for you. It comes with over 100 built‑in metrics and tests, covering everything from data quality to model drift, from retrieval quality in RAG to hallucination detection in LLMs. If you want to customise, you can go from Python functions to LLM judges – it's up to you.
I initially used it for data drift detection, and the results were quite solid. Then I discovered it could even evaluate RAG systems – measuring retrieval quality and generation quality separately, so when something goes wrong, you can pinpoint exactly where the problem lies: whether it didn't find the right information, or it found it but used it incorrectly. That's incredibly useful because RAG pipelines are long, and when something breaks, you have no idea who to blame.
What impressed me most...
What impressed me most is that it's completely open‑source, Apache 2.0. It has 7500+ stars on GitHub and over 40 million downloads. Companies like DeepL, Wise, and Plaid are using it. One user described it as a "Swiss Army knife," saying they use it more often than expected. That's high praise – after all, no one praises a screwdriver every day.
Evidently isn't without its downsides. The documentation can sometimes be too detailed, making it a chore to navigate. Also, if you're doing pure real‑time online monitoring, you might need to pair it with other tools. But it does solve my core pain point: it lets me know what my AI is actually doing.
If you're also building AI products, my advice is: don't wait for your users to find the bugs for you. Spend half an hour going through Evidently's quickstart to see how your model performs on real data. Even if it's just a one‑off evaluation report, it's still better than flying blind.
After all, it's not sc...
After all, it's not scary that AI makes mistakes – what's scary is not knowing that it's making them.