It Learns How You Classify

Fixing a category once is fine. Fixing the same one for the tenth time means the product isn't listening.

Every business has its own logic. Your Friday Careem is a client meeting, not personal transport. That supplier everyone else files under "General" is, for you, clearly "Cost of Goods Sold". A default rulebook will never capture that — and honestly, it shouldn't try.

So we made SnapLedger listen instead.

Your corrections don't disappear

From now on, every time you reclassify an entry — change a subcategory and hit save — we record it as a correction event: what the entry was, which way the money moved, what it was, and what you changed it to.

On its own, that's just a log. The interesting part is what happens next.

Corrections become rules — written like ours

When you next import documents, an AI job reads your recent corrections and distills them into a small set of personal rules, written in exactly the same format as our built-in classification rules. "Description mentions X → subcategory Y" — but yours.

At classification time, the model sees both: the default rulebook first, then your rules, with an explicit instruction — when yours match, they win. One model, one pass, your judgment layered on top of ours.

New corrections keep arriving? The rules regenerate. Teach it once, and it remembers; teach it something new, and it updates.

The line we won't cross

Here's what we deliberately did not build: learning across users. Your rules are derived from your books, stored under your account, and applied only to your classification. Nobody else's data shapes them, and yours shapes nobody else's.

And the discipline from before still holds: the pipeline stays AI-driven with the same human-review rules — no hidden lookup tables, no silent overrides.

A ledger should work the way you think, not the other way around. Correct it once — it remembers.