NumbyAI: Bank Format Handling & FAQ
If you have done enough CSV work, you know the issue is never "AI categorization." It is always imports. Most finance apps assume bank exports are clean. Then your CSV lands with metadata rows, ambiguous date formats, mixed decimal styles, and split debit/credit columns.
What NumbyAI detects
- •Metadata preamble rows (scans until it finds date + amount)
- •Column roles using density signals, not hardcoded bank names
- •9 date format patterns including common US/EU variants
- •Number formats: EU style (1.234,56) and US style (1,234.56)
- •Currency from symbols and ISO codes
- •Inflow/outflow split columns and balance columns
- •Delimiter style including tab and pipe formats
Supported bank formats
Tested: Chase, Bank of America, Wells Fargo, Barclays, HSBC, ING, Sparkasse, BNP Paribas, UBS, Revolut, Santander — plus generic cases with metadata rows.
FAQ
Is this a cloud app? No cloud upload by default. The project is local-first: no cloud uploads, no data sharing, no telemetry. Default model path is local Ollama.
Can it handle everything? Not literally, but it handles the messy cases people trip on: missing metadata rows, multiple delimiters, bank-specific debit-credit patterns, unusual date and number styles.
What happens when confidence is low? LLM fills the gap using structured prompts and sample rows. Uncertain transactions land in a review queue, not silently categorized.
My bank is not listed. What now? Start with a sample import to identify where parser confidence drops, then add coverage via rules and fixtures. The "Add new bank format" contribution path is documented.
How do I avoid accuracy errors? Use the review queue as a quality gate. Trust bulk automation on normal rows, spend one pass on review conflicts, then convert repeatable mistakes into rules.