Anomaly Detection

Anomaly detection is Geekonomics's safety net — a daily scan that flags transactions that don't look right so you can fix them before they end up in a financial statement.

What It Catches

Five detectors run together, each looking for a different kind of problem:

• Duplicate transactions. Same vendor, same amount, same date (or within a day). Catches double-imports from Plaid, accidental re-entries, and sync bugs.
• Spending spikes. A category whose monthly total jumps more than a configurable multiple of its trailing average. Often a sign a one-off got miscategorized.
• Missing receipts. Expense transactions above a dollar threshold with no receipt attached. Useful for audit prep.
• Old uncategorized transactions. Anything synced in, not yet categorized, and older than a configurable number of days.
• LLM outliers. Claude reviews recent transactions in context and flags entries whose description doesn't match the category, the amount, or the vendor in a way the rules-based detectors would miss.

When It Runs

A cron kicks the scan every morning at 4 AM. Findings land on the Anomalies page in the sidebar with the most recent at the top.

You can also click Deep Scan to run the LLM detector on demand — useful right before close or before sending books to an accountant. Deep Scan uses your AI token budget.

Reviewing Findings

Each finding shows what was flagged, why, and links straight to the underlying transaction. Two actions:

• Resolve — you fixed the issue (recategorized, deleted the duplicate, attached the receipt). Removes the finding.
• Dismiss — the finding is wrong, or it's a one-off you don't care about. Removes the finding and tells future scans to ignore the same pattern on that transaction.

Resolving and dismissing are both logged in the audit trail so there's a record of who looked at what.

What's next

AI Assistant — for proactive bookkeeping help.