Same here... back in 2011 or so. We needed something much more performant than WURFL. My efforts eventually became a feature/product at Akamai known as "Edge Device Characterization" (EDC) using algorithms not dissimilar to how LLMs are trained today.

I can't speak to how good the actual product is today (or even when it launched, but that's a whole 'nother story), but during development it was capable of processing 100K RPS in a footprint of ~30MB RAM with ~98% accuracy compared to WURFL as a baseline.