Saved me hours! Super easy to use app.
I was manually entering receipts into my spreadsheet like I was from the 1900's. I decided to buy this deal and it literally has paid for itself 5 times over. My box of receipts I was procrastinating about did take me 2 hours to record, but it would have taken 10 manually.
Only negative is sometimes the app has an error message and you have to reenter data from a receipt or take the photo again. It added about 10 seconds to my workflow. It happened on about 2% of my receipts.
The other thing to watch is date format. Being in europe the date format is opposite from the USA. Maybe there is a setting I missed to change this? Perhaps the AI can be trained to see the location of the store and read the date accordingly.
The receipt reader also sometimes grabbed a "total" from the VAT tax chart instead of the actual total. It was an easy fix to manually enter the correct total. I also would try to not take a photo of the VAT chart (which is usually just below official totals).
Antti_SparkReceipt
May 9, 2024Hi Calvin,
Thank you for the great and in-depth review!
Sorry to hear about the error messages you've encountered. But the good thing is that these should all be catched in the error reporting system. I periodically have bug fixing sessions in between bigger updates to address any reports that arise or that escalate.
And as you said, date formats can be difficult. There is a setting in the web application to set the preferred date locale, and this setting is heavily preferred when dealing with ambiguous dates (e.g. 3/1/2024 vs 1/3/2024). In the mobile app this setting is automatic based on the phone's region setting. Currently the AI-system for dates is two-phased, where some preliminary filtering and parsing is attempted with traditional algorithms before using the AI. This is because the model has not been powerful enough to do the reasoning based on context alone, resulting in common misdeductions of date formats. I will soon move this process to a better model since the AI pricing is (fortunately) coming down, fast. It should improve the results further.
VATs that are displayed in charts is a known weak point of the current algorithm. Especially if there is e.g. a table of 2-3 VAT tiers. I know this first hand since this is a format they use for receipts in Finland. And with paper receipt OCR the layout is not always fully preserved for the AI to piece everything "back together" in the table. I have some future plans to make this better, but it has proven to be one of the more difficult problems.
Thanks,
Antti