The Personal Audit Engine

Imagine a personal finance tool that doesn't just track your spending but makes you think about each purchase as it happens. The Personal Audit Engine is a mobile-first app that replaces the spreadsheet with a single daily 'audit log' where you enter each transaction as a short tweet-like entry, and the app assigns it to a virtual envelope based on machine-learned categories. The catch: you have to 'win' the envelope allocation by correctly guessing the category before the app reveals it. This game-like mechanism trains your financial intuition over time.

The engine is built entirely by one person, a solo developer who is also a behavioral finance enthusiast. The architecture is deliberately simple: a mobile app with a local-first database, syncing to a single-user server only for backup and model training. The whole system is designed to be maintainable by one person, with no external dependencies except the user's own discipline.

The core loop is daily: user records each purchase with a short note (e.g., "coffee shop" or "Netflix"), the app shows its best guess for the envelope (Food, Entertainment, etc.), and the user confirms or corrects it. Each correct guess adds to a streak, building a personal 'accuracy score.' Over time, the model learns from corrections, making better predictions. No bank connections, no automatic imports—just pure, reflective logging with a playful edge.

Traditional budgeting tools demand constant data entry and categorization, which feels like a chore. Even automated tools remove the reflection that leads to better habits. The Personal Audit Engine scratches the itch of wanting a tool that is both effortless and educational, turning each transaction into a low-stakes game that reinforces financial awareness.

• Build a mobile app that allows users to log transactions in under 10 seconds.
• Implement a machine learning model that accurately predicts envelope categories from raw text.
• Create a 'guessing game' UI that rewards correct category predictions with points and streak achievements.
• Achieve 95% accuracy in category prediction after 100 transactions per user.
• Run the entire backend on a single $5/month server.
• Maintain a zero-debt personal finance philosophy in the app's design.

• The app will not connect to any bank accounts or pull data automatically.
• It will not produce fancy charts or reports—only the daily audit log.
• It will not support shared budgeting or family accounts.
• It will not attempt to replace a full accounting system; it's for personal cash flow awareness only.

Flutter for cross-platform mobile app (Dart), SQLite for local storage with Firebase Cloud Firestore sync for backup, Python (Flask) for a lightweight ML API using scikit-learn for text classification, deployed on a DigitalOcean droplet. No external accounting APIs. The ML model uses TF-IDF vectors and a simple logistic regression, trained on the user's own transaction history.

The app follows a single-user, local-first architecture. The core data entity is the 'Entry' – a timestamp, a short description (text), an amount, and a true envelope label. The user enters an entry, and before saving, the app shows a guessed envelope from the local ML model. The user can accept or change it. Each entry becomes a training sample for future guesses. The ML model is trained locally on the device using TensorFlow Lite (though initially just a simple Bayesian classifier). The sync to the cloud is a JSON dump of all entries, encrypted. The server is only used for model updates (a periodic batch re-training from all user data, but since it's single user, it's optional).

• User abandonment if the game becomes tedious.
• ML model accuracy may plateau without manual corrections.
• Single point of failure: if the solo developer gets ill, the app stops receiving updates.
• Privacy concerns if cloud sync is not properly encrypted.

• Should the guessing game allow 'skip' without penalty to reduce friction?
• How to handle recurring transactions (e.g., Netflix) that the user might want to categorize automatically?
• Is the local ML model sufficient, or is cloud training needed?
• What reward system works best for long-term engagement?

The solo developer started building the prototype in Flutter but got overwhelmed by the ML integration and never released even an alpha. Life happened—a job change and a move—and the project languished on a shelf of half-finished directories.

The project was initially conceived as a way to practice full-stack development while solving a real personal pain point. It remains a proof of concept in the developer's heart.