I have been poking around the community this weekend, and it looks like the most interesting use of this receipt scanning mechanism would be to take the OCR and parse out the vendor and total price. Then auto-matching a cloud-based receipt URL to a Receipt column on the transactions tab.
The first bit is pretty easy to adapt my current setup to. As far as doing the auto-matching, I got Claude to generate some app script that relies on confidence levels around [sub]string matching in the transaction description and the OCR detected vendor. I am wary about inserting Claude’s GAS into Tiller, but it seems like whatever mechanism AutoCat uses could be adapted to run through the [sub]string matching.
Storage can be on AWS S3, which provides for private signed links, in my understanding of the privacy-first Tiller Way. I would bet other hosting providers do something similar. I don’t think this solution would be very expensive to implement and run, unless you really wanted to stay off commercial LLMs and were using local LLMs. The GPU costs can stack up quickly. But maybe it’s manageable at Tiller’s scale.
What think you Tiller folks? Is there enough plumbing at this point to take this on as a feature?
Update My progress on auto-matching is here: Best practice for Google Application Scripting in Tiller - #2 by BryanS. It relies on a custom function and an ARRAYFORMULA(). The error handling could be better, the substring matching can be extended, and it is still a bit slow. Might time out on seriously large sheets.