Your card statement knows more than it says.
We infer the items behind each charge — from the amount, the merchant, and your own habits. No receipt photos. No manual entry.
Arithmetic before AI
Tax reversal, tip decomposition, and exact-sum search over real menu prices are done with exact integer arithmetic. A language model ranks which baskets are plausible for you — it is never asked to do the math.
Honest about uncertainty
Every answer is a probability distribution, not a single confident guess. When thousands of item combinations fit a charge — a big grocery run, say — the app tells you exactly that, along with the items it can still vouch for, instead of inventing a receipt.
It learns from you, one tap at a time
Confirm or correct a single prediction and every future inference at that merchant gets sharper. Your corrections build a personal prior — the app learns your usual order, not the average customer's.
Private by architecture
The inference engine only ever sees pseudonymized data — an opaque identifier, never your name, email, or account number. Model prompts carry the merchant, the amount, and candidate items; they never carry who you are. We only work with model vendors under terms that forbid training on submitted data. And we say “pseudonymized,” not “anonymous,” because purchase histories are identifying by nature and we would rather be precise than reassuring.