ai-bookkeeper reads receipts into merchant, date and total. On 626 real Malaysian receipt scans (SROIE) run through tesseract, the merchant came out exactly right 33.1% of the time. Three small changes lift that to 36.6%: skip lines that are mostly symbols, GST registration numbers and state names; strip stray marks before the name; and read the engine's SON BHD as SDN BHD.
The first attempt, filtering any line with few letters, looked better on OCR text but cost 7.6 points on clean text, because names like MR D.I.Y. and 99 SPEED MART have dots and digits. Counting symbol characters instead kept clean-text accuracy flat (63.7% to 63.6%). Most remaining misses have no readable shop-name line at all. Phone photos are still untested. Estimates only; check amounts before filing.