Update: Measuring before improving
One lesson from building Textura is that improving AI isn't just about changing prompts, it's about measuring quality consistently.
Over the past few days I've built a repeatable evaluation pipeline using manually transcribed ground-truth documents. Current results: • Personal journals / letters: WER = 0.649 ± 0.031 (n=7) • Official documents: WER = 0.655 ± 0.032 (n=7)
The next step is expanding the benchmark with additional historical documents from different writers before experimenting with image preprocessing, multi-pass transcription and ensemble approaches.
Finally, our first Instagram Reel is now live, sharing the story behind why Textura exists in the first place.
https://www.instagram.com/reel/DZ9iso5oLsr/?igsh=MWppMnpzNng2djJ1eQ==
Building continues.
Log in or sign up for Devpost to join the conversation.