
The model inside Spends: a half-billion-parameter language model, fine-tuned to read Indian bank SMS and turn each one into a transaction: how much, which way, to whom. It gets 98.5% of messages fully right, against 93.5% for my hand-written rules, and it is small enough to run on the phone, so nothing leaves it. The playground shows its answer token by token, with how sure it was of each one.





