The product
Gnarly Budget is Gnar.li’s own budgeting application, in development since 2024 and deployed with trial users. It is built around a weekly routine that makes budgeting stick: import often, review monthly, check cash flow weekly.
The challenge
People who budget seriously often live in spreadsheets because budgeting apps either want direct access to their bank or don’t match how money actually arrives. Gnarly Budget had to be as flexible as a spreadsheet, plan around real paydays, and keep financial data private, with no bank connections, no ads, and no data selling.
What I built
- Budgets anchored to payday. Budget periods follow the payroll schedule, line items edit like a spreadsheet, and each new period starts as a clone of the last.
- Imports instead of bank connections. CSV or OFX from any bank, or pasted from a spreadsheet. Duplicates are skipped, pending and upcoming transactions are matched, and categories are assigned automatically.
- Forecasting. Recurring templates schedule bills and paychecks, so the cash flow forecast covers the whole period, not just today.
- Balances you can trust. Every account can be verified against its statement closing balance, so discrepancies are caught before they compound.
- Ledger detail. Split transactions, linked transfers between accounts, saved snapshots for flexible spending, and reports for spending, balance sheet, cash flow, and custom labels with CSV export.
- Planning ahead. Draft up to three budget periods ahead and track a financial independence target with compound growth projections.
- Data control. Account sharing with a partner, full export and deletion, and session review.
Where AI fits
AI plays a narrow role: triaging support tickets, answering questions in help chat from the app’s documentation, and summarizing the monthly budget review. It runs on self-hosted Ollama models, with no third-party LLM API calls.
The numbers never come from the model. Financial calculations live in tested Ruby services, model output is validated for numeric grounding and prohibited advice, and every review card has a template fallback if generation fails.
Lessons
Let tested code own the numbers. Software that people trust with their money has to be right first, and helpful second.