show and tell: i shipped BhashaOCR, a Bahasa Melayu + Jawi OCR pipeline, in 6 days, end to end. 1.2k receipts/day on the Bungkus test rig, 91% accuracy on Bahasa item names, MIT licensed.
the stack: Tesseract for line detection, a fine-tuned BERT-Mini for Bahasa item-name post-processing, and a custom CORS proxy because the public OCR APIs all wanted RM 1,200/month for a "starter" tier.
the bungkus circle reviewed the PR. if you want to see the code, the repo is on github.
— @rendra.fkaas · 28 Apr 2026
<3 replies>
@hafiz.fskm28 Apr, 12:12
this is exactly the kind of work that should be in the UTHM Forge showcase. 6 days from "i have an idea" to "1.2k receipts/day on the test rig" is a real shipping velocity. +1.
@saifuddin.fke29 Apr, 00:00
the CORS-proxy trick is the right move. the public APIs price themselves out of the indie / FYP-grant market on purpose. building the proxy + the BERT post-processor in 6 days is the kind of leverage a small community has over a large company.
@ameera.fstmi29 Apr, 07:45
linking this from the showcase. deserved.