From Prompt to Play Store: Building a Localised Money App Mostly by Prompting AI Agents

Ledgr is a shipped Android expense tracker for Bangladesh that turns bank and mobile-wallet SMS into a categorized spending history with zero manual entry, and I planned and built almost all of it by driving AI coding agents rather than hand-writing the codebase.

The app now spans 86 Kotlin files, an 11-table Room database at schema version 14, and 68 SMS parser templates across eight local financial senders (SCB, EBL, Trust Bank, MTB, BRAC Bank, City Bank, bKash, and Nagad). It is live on the Play Store.

For the demo I will show two things, both technical, no slides:

The AI development workflow that produced it. I will open the actual living context documents I feed to every agent session (a single tech_details.md that describes the whole architecture, schema, DAOs, and gotchas, plus a product_features.md), the spec files under PLANS/ that I write before each feature, and how I keep those documents in sync with git so the next agent session starts from ground truth instead of re-deriving the codebase. I will walk one real feature from spec to shipped code.
The running app in Demo Mode. I will show the localised AI features that came out of that workflow: voice and text transaction capture that parses Bangla, English, or mixed input such as “kalke cha 60 taka” into a structured transaction, the on-device SMS parser turning raw bank messages into transactions, and the AI Spend Report and insights generated from local aggregates.