Inference runs on your phone. No cloud AI API, ever.

AI that never sees your data leave your device.

Ask questions about your money and get answers read from your own records, on your phone. A cloud chatbot has to send those records to a server before it can answer. This one has no server to send them to.

Local LLM inference

PocketVault Finance runs Gemma language models entirely on your device via flutter_gemma. Choose from multiple models based on your device's capabilities, or use Apple Foundation Models on supported iOS devices with no download at all.

  • Choose from Gemma 4 E2B, Gemma 4 E4B, or lightweight Qwen 0.6B
  • Apple Foundation Models on iOS 26+, no download needed
  • No internet required for AI features

LLM Engine

Runtime Gemma (flutter_gemma)
Models Gemma 4 E2B | E4B | Qwen 0.6B
Processing 100% on-device
No data leaves your device
"What do I spend on coffee?" Ask

Records it pulled in

Starbucks 98%
Blue Bottle 95%
Dunkin' 91%

How it finds the right records

A question does not go to the model with your whole ledger attached. The records that relate to it are retrieved first, using EmbeddingGemma 300M embeddings computed on the phone, so a question about coffee reaches a café that never uses the word.

This retrieval runs inside Ask. It is not a separate search box — the app's transaction search is its own, and matches on text.

Categorising is not the model's job

Worth being plain about, because most apps imply otherwise: the language model does not categorise your transactions. Rules do, in a fixed order — anything the bank's own message stated, then how you categorised that merchant before, then a merchant list that ships with the app.

That is a deliberate choice rather than a missing feature. A rule you correct once stays corrected, and you can see why it decided what it did. A model's guess offers neither.

Your history first

However you filed that merchant last time

Then the catalogue

A shipped merchant list, for names you have not seen

Correct it once

And that merchant stays corrected

Natural language queries

Ask "How much did I spend on food last month?" and get streaming answers powered by on-device RAG that pulls context from your transaction history.

How much did I spend on food last month?
$342.50 across 23 transactions. That's 15% of your total spending.
Powered by on-device RAG

It looks things up. It can’t change anything.

The AI answers by calling the app’s own calculations, the same ones behind its screens, and reading what they return. Every tool it has is read-only: there is no way for it to add, edit or delete a record.

  • Looks up: spending totals, a month against your usual, transactions, bills coming due, budgets, currency conversion
  • With Premium: subscription review, loan payoff scenarios, and where your net worth is heading
  • Checked figures: a number the app didn’t calculate is held back rather than shown

AI Chat

Which bills are still due this month?
Two: rent, $1,450.00 on the 28th, and internet, $65.00 on the 30th.
Bills
Read-only. It answers from your records and never edits them.

Why it runs on the phone

An assistant that answers from your records has to read them, and where it reads them decides who else could. Somewhere in a cloud app's stack, those records sit on a server. Here they stay on the device they were already on, so answering costs you nothing in exposure.

That is the whole claim, and it is worth being precise about its edges: the answers are not better than a large cloud model's, and the model does not get sharper the longer you use it. What changes is that there is more of your own record for it to read.

Ask is free. Premium is the analysis that needs no model at all: recurring charges you had forgotten, each subscription priced by the year, and extra loan payments modelled against the interest they save.

Keep your finances to yourself.

Every feature you need is free, and there's no account to create. Download it and your data stays on your device.