An AI expense tracker that never sees your expenses

The problem with AI in money apps

Almost every app that added an AI assistant in the last two years did it the same way: ship your transactions to a language model and print what comes back. That design has two flaws, and both of them are structural rather than fixable with a better prompt.

• Your ledger becomes someone else's input. Not a summary — the rows. Where you were, what you paid for, who you paid. Whatever the retention policy says today, the copy exists.
• Language models produce plausible text, not arithmetic. Ask one how much you spent on food and it will answer in the shape of a correct answer. Whether the figure is right depends on how well the numbers survived the trip through a probabilistic model. For a chart that is a bug; for "can I afford this" it is a decision made on a made-up number.

Neither problem is anyone's incompetence. They are what you get when the answering machinery is a text predictor that lives on a server.

How Klera does it instead

Klera splits the job in two — understanding the question, and answering it — and refuses to let a model near the second half.

1. Your question is matched, not generated The app holds a catalogue of question shapes across 22 families — affordability, spending, budgets, goals, udhaar, bills, loans, subscriptions, anomalies, habits and more. Your words are normalised (English and Hinglish share the same catalogue) and scored against it. One clear winner proceeds; a close tie asks you which you meant; anything below the floor gets an honest refusal.
2. Missing details are asked for, not assumed "Can I afford a phone?" has no price in it, so Klera asks for one instead of inventing a number. Amounts written the way people actually write them — 30k, 1.5 lakh, 2 crore — are read exactly.
3. The answer is computed from your ledger One snapshot of your own data — balances, run-rate, committed outflows, budgets, goals, IOUs, upcoming bills — feeds a plain function that does arithmetic. Money is handled in whole paise end to end. The same code that draws your charts produces the number in the chat.
4. You see the working Answers arrive as a card: the verdict, the sentence, and the figures behind it. Ask the same thing tomorrow with the same ledger and you get the same answer — which sounds obvious until you have used an assistant where it is not true.

What you can actually ask

Real question shapes from the shipped catalogue, in English or Hinglish, typed or spoken.

And it writes as well as reads: type "200 chai, 300 fuel, 200 movie" and Klera logs three expenses, categorises each, and shows them back as cards you can tap to edit. Anything phrased as a question goes to the answering engine instead — the two paths never collide.

On-device AI vs cloud AI, honestly

The trade is real and runs the other way too: a language model will understand a sentence Klera has never seen, and can write you a paragraph about your finances. Klera answers a known set of questions exactly, and says so when yours is not one of them. For money, we think that is the right side of the trade — but it is a trade, not a free win.

The privacy guarantees, specifically

No LLM, by design

Klera does not call OpenAI, Anthropic, Google or anyone else, and does not bundle a local model. The understanding layer is a catalogue of question shapes; the answering layer is arithmetic. It is an architecture decision, written down as ADR 0050 in the app repo, not a setting we could quietly flip.

No network call to answer you

There is no request to fail, log or leak, and you can verify it the way you verify the rest of the app: turn on airplane mode and ask it a question. One honest exception — dictation. Klera asks your phone to transcribe on-device, but without an offline language pack Android may route the audio through Google's recogniser. That is the phone's own speech engine rather than anything of ours, Klera never records or keeps the audio, and typing avoids it entirely.

Nothing is stored

The conversation lives in memory while the screen is open and dies with it. Questions and answers are financial content, so they are never written to disk, never logged, and never counted in any analytics — there is no Klera analytics to count them in.

It shows the figures it used

Answers come with the numbers behind them — your liquid balance, your run-rate, your safety buffer — so you can check the working rather than trust a tone of voice.

It refuses rather than guesses

Below its confidence threshold the answer is "I can't answer that yet", plus examples of what it can. A wrong number about your money is worse than no number.

The rest of the app is built the same way — every feature works offline, there is no account, and the database is AES-256 encrypted on your device. The AI did not get an exception.

What it will not do

Stated plainly, because an assistant that oversells itself is worse than one that does less.

• It answers question shapes it knows, not any sentence ever written. Coverage is measured against a bank of 1,316 real phrasings and stands at 94% — answered outright, or after it asks which of two readings you meant. The remaining 6% get an honest refusal instead of a guess.
• It will not write you an essay, plan your taxes, or recommend a mutual fund. Tax and insurance questions are routed to a deliberate "I don't answer that" rather than an opinion.
• It reads your Klera ledger. If you have not logged much yet, it says so and asks for a month or two of history rather than extrapolating from nothing.
• It is a personal-finance assistant, not a general chatbot. Ask it about the weather and it will tell you it cannot help.

When Klera speaks first

The same on-device engine can raise its hand without being asked — a weekly money story, a bill due tomorrow, a budget nearing its limit, an account running low.

• Every type is off by default. Nothing arrives until you switch it on, type by type, with one master switch above them all.
• Worked out on your phone. No server decides you are overspending, because no server knows.
• Quiet by design. Nothing between 10 PM and 8 AM, and amounts stay out of the notification shade unless you choose otherwise — a locked phone should not leak your balance.
• No urgency theatre. The tone rules are the same ones the udhaar reminders follow: never shame, never panic.

AI expense tracker FAQ

Read next

• Why an AI expense tracker could beat every other expense tracker
• Offline expense tracker apps, explained
• Should you give an expense app your bank SMS?
• Best expense tracker apps compared
• Every network call Klera makes