Skip to content
Jai Jhamb
All work
Fintech · 2026 · Personal project · In progress

Wallet

Personal finance system with an integrated AI layer.

TypeScript · Expo / React Native · Kotlin · SQLCipher

01Problem

A mobile app for personal finance. Money is recorded in a ledger on the device, and an AI layer sits beside it — never inside it.

I wanted something more serious than a basic expense tracker: to understand personal finance as a system — how transactions, balances and budgets should actually fit together — and then see where AI could help without being allowed to decide anything on its own.

02My role

Solo

03What I built

  1. You

    Record, confirm, decide

    Every change is a user action

  2. Finance system

    Double-entry ledger · deterministic engines · encrypted on the device

    Read-only: the AI never writes

  3. AI layer

    Explains, suggests, abstains when unsure

The design rule: the AI can read the numbers and make suggestions. Only a user action, recorded in the ledger, changes them.
  • A double-entry ledger: append-only, corrected only by reversals, with time-sortable IDs.
  • Tested, deterministic engines for cash-flow projection and spending analysis.
  • Offline-first: data lives on the device, in an encrypted SQLite database.
  • An Android payment-capture module in Kotlin: incoming bank SMS are matched against templates and held in an encrypted staging queue.

In the lab

  • On-device categorisationexploring

    Not built yet. Today’s categoriser is a rule cascade that abstains when unsure.

  • Forecasts as rangesexploring

    A proposal only. The deterministic forecast stays the authority.

04Engineering decisions

    • Option A

      A mutable balance column

    • Option BChosen

      A double-entry ledger

    Trade-off
    A balance column is simpler to write. A ledger costs more code and storage, but every number can be rebuilt and checked.
    Decision
    Append-only double-entry ledger. Corrections are reversals, and the books are checked after every change.
    Why
    Every balance traces back to source rows, so it can’t silently drift — and a ledger is very hard to retrofit later.

    • Option A

      AI computes the figures

    • Option BChosen

      AI only explains them

    Trade-off
    Letting a model compute is faster to build. Keeping it out of the arithmetic means more layers, but the numbers stay exact.
    Decision
    The AI layer computes nothing. It can suggest, but it never writes to the ledger without a user action.
    Why
    Increasing model intelligence must never increase model authority.

    • Option A

      Make its best guess

    • Option BChosen

      Abstain

    Trade-off
    Guessing always gives an answer. Abstaining sometimes gives none.
    Decision
    Category suggestions abstain, forecasts refuse to invent an exchange rate, and safe-to-spend has explicit refusal states.
    Why
    In finance, a wrong number costs more than a missing one.

    • Option A

      The whole SMS inbox

    • Option BChosen

      Only new messages, as they arrive

    Trade-off
    Inbox access allows backfilling history. New-only access sees less, but asks for far less.
    Decision
    Receive new SMS only, never read the inbox, and process everything on the device.
    Why
    Least privilege — and it is the strongest argument in a Play Store review.

    • Option A

      Real time and real randomness

    • Option BChosen

      Injected time, seeded randomness

    Trade-off
    Real inputs are simpler to wire. Injected ones take discipline, but results reproduce exactly.
    Decision
    The forecasting engine guarantees no clock and no randomness.
    Why
    A test that can’t reproduce a number can’t protect it.

05What I learned

I built two complete engines before any screen could reach them. Breadth reads like progress; users only get value from what they can reach.

06Links

No public repository or demo.