No OCR actually - just a vision model doing the recognition, one
prompt, JSON out (Qwen3.6-27B on Groq).
I didn't chase some clever accuracy trick, because like you said, a
photo alone can only get you so far. The design bet was different:
make it fast and make every number editable, so the friction stays
low and you fix the guess instead of trusting it.
It nails simple plates and gets confused by messy/mixed ones — that's
the part I'm still working on.
I didn't chase some clever accuracy trick, because like you said, a photo alone can only get you so far. The design bet was different: make it fast and make every number editable, so the friction stays low and you fix the guess instead of trusting it.
It nails simple plates and gets confused by messy/mixed ones — that's the part I'm still working on.