Usually is an open-source personal finance application designed to provide contextual spending insights without requiring manual spreadsheet entry or sharing raw banking logs with cloud aggregators, according to project documentation and developer posts. Conceived during the Hacktoberfest Open-Source AI Challenge, the project draws inspiration from operating system telemetry screens that compare daily resource consumption against historical baselines rather than displaying isolated monetary figures.
Deterministic Financial Accounting
The application relies on a strict architectural boundary that separates pure mathematical calculations from language model assistance. According to the project architecture, all core accounting tasks—such as cash-flow sums, income tracking, living expenses, transfer categorizations, and spending-pace benchmarks—are executed entirely through deterministic Python logic and SQLite database queries. By isolating arithmetic from probabilistic text processing, the system prevents large language models from hallucinating balances, inventing totals, or misinterpreting sensitive financial records. Transaction parsing normalizes standard CSV, TSV, and PhonePe exports, applying content-addressable SHA-256 deduplication before records enter the private workspace ledger.
Local-First AI Merchant Classification
When the application encounters unfamiliar or ambiguous merchant descriptions, it utilizes local open-weight models to suggest standard categories without transmitting private transaction records over external networks. In development setups, Google’s Gemma 2B model runs locally via Ollama to evaluate messy transaction strings and propose one of twelve canonical expense categories. However, the system enforces a human-in-the-loop requirement where AI suggestions are treated strictly as candidate proposals requiring explicit user review. Once a user approves or overrides a suggestion, the decision is permanently recorded in the local SQLite database, allowing future imports of the same merchant to bypass model inference entirely.
Why Bounded Local AI Matters for Financial Software
The design choices behind Usually highlight a critical architectural lesson for developers building software that handles private personal data. Traditional financial technology tools often force an undesirable compromise between high-friction manual spreadsheets and invasive cloud aggregators that harvest credentials and transaction logs. More recently, conversational AI financial bots have introduced new failure modes, including arithmetic hallucinations and privacy risks associated with routing sensitive figures to third-party model APIs. Usually addresses these vulnerabilities by establishing a zero-trust boundary around probabilistic components. By restricting Gemma 2B solely to messy string interpretation while locking core calculations inside deterministic Python modules, the project demonstrates how developers can integrate modern language models without sacrificing accountability or predictability. Furthermore, running models locally via Ollama ensures that users retain control over their inference pipeline and data flow. This pattern matters because it decouples intelligence features from cloud dependencies, offering a viable blueprint for building privacy-conscious applications where users remain the ultimate authority over automated decisions.
Architecture and Verification
The codebase is structured around a modern web technology stack, featuring a FastAPI backend paired with a Next.js App Router frontend. Personal data is segregated into individual SQLite workspaces operating in Write-Ahead Logging mode to ensure physical workspace isolation and straightforward session management. To validate system reliability, the repository includes comprehensive automated testing suites covering offline backend unit tests, frontend component checks, and integration tests for local model workflows.
Editorial graphic and headline text by Signal Titans.
Source: I Wanted an iPhone-Style “More Than Usual” Alert for My Money. So I Built Usually. · GitHub – debashis-sharingan-21/usually: Your money has habits. Usually knows them. · GitHub
SIGNAL TITANS
Leave a Reply