Act II · Money · Step 04 of 10
Bank and UPI SMS, read on your phone, into categories and a monthly pace. Where the map ends, it says so on the screen.
A confident pie chart never saw the ₹300 you paid in cash. It is wrong in the direction that makes you feel fine.
The honest version is less pretty: here is what I could see, here is the total I can vouch for, here is where my vision ends. You can reason with that, because you know what it leaves out.
Tracked spend · March
₹18,775
Cash and some cards are invisible here. ARTH does not guess the gap.
Every number here says tracked, and means it.
A figure you can vouch for beats a figure that looks complete.
Most of this never leaves your phone. What does leaves stripped to almost nothing.
Optional, revocable. Withhold it and the spend map stays empty rather than falling back on a worse guess.
Rules pick out bank and UPI transaction SMS. Personal messages, OTPs and promotions are skipped on-device.
Amount, direction, date and merchant, all read locally. Most everyday merchants finish here and travel nowhere.
One item per payee, not per transaction - so a hundred food orders cost one lookup. Digits and amounts stripped first.
A category from a fixed list, with a confidence rating. Nothing is kept, and a hard ceiling bounds the spend.
Recategorise anything and the label sticks as an override. The next scan leaves it alone.
Twenty-one categories, kept small. A taxonomy nobody can hold in their head is one nobody corrects.
The design decision is not that ARTH uses a model. It is how little the model sees, and how rarely it is asked.
OpenAI
Merchant categorisation
On-device rules resolve most transaction SMS on their own. What is left is one entry per distinct merchant - redacted, with long digit runs and exact amounts stripped - sent for categorisation. A cheap model does the labelling; only genuine disagreement escalates to a larger one.
Never sees your full inbox, personal messages, or account numbers. Your correction always overrides the label.
Vendor names and logos belong to their owners and are shown here only to describe how ARTH is built. Tax figures are never produced by a language model - those come from a deterministic, versioned rule engine.
Held on purpose. Not a gap waiting for a roadmap.
Your bank login stays with your bank. Any account access will go through a regulated Account Aggregator, with your consent.
Personal messages, promotions and OTPs stay where they are - unread, unstored, unsent.
Tracked spend is labelled tracked. The money ARTH could not see is left as a stated gap, not an estimate.
Your label is the final label. No later scan, and no second opinion from a model, overrides it.
Controlled Android beta · Tax Year 2026-27 · Income-tax Act, 2025
ARTH shows where your money really goes, without rounding it into neat categories.