We've been building GenAI applications using open-source models from Hugging Face. Earlier, it was a real struggle to feel confident that we had selected a performant model for our apps, as our approach and criteria for model selection were quite ad hoc. Future AGI's evals framework was intuitive and easy to hook up with our codebase. Not only did it give us a methodical approach to GenAI model selection, but it also allowed us to systematically improve our GenAI apps by enhancing model performance. The results were evident in a double-digit improvement in our online product metrics.

The hallucination and toxicity tests caught issues even our reviewers missed.

It caught a hallucinated answer in our chatbot during testing before it reached production. That one moment paid for the whole tool.