Doze: teach it once. Let the AI sleep.

Description and use

Doze trains a local Gemma model once to map columns between payment and bank exports, saves the mapping as JSON, and reuses it without further inference, performing exact arithmetic checks and flagging discrepancies.

Problem

Recurring AI workflows repeatedly call the model to reconcile payment exports with bank exports, wasting inference resources and time.

Outcome

Doze achieves 52‑week replay with 41 clean, 11 held for review, and 19 explicit exceptions, all without any model calls during replay. Eight fixed mapping tests pass, and the system flags a one‑cent discrepancy as a mismatch.

Workflow

• Teach the model once with headers and human‑approved explanations
• Gemma returns a mapping or refusal
• Human reviews and approves the mapping
• Save the mapping as JSON
• For each new export, load the approved mapping
• Python reconciles payments and bank data using exact arithmetic
• Flag discrepancies or approve clean weeks
• Generate audit ZIP with reconciliation CSV, audit JSON, mapping, and hashes
• Optionally run via CLI without browser or inference

Evidence

Doze refuses to close a week that is one cent short. It can do that with inference disabled.

Source

The chore is reconciling payment exports against bank exports.

Source

Doze uses local Gemma to learn the column mapping, then saves it as a small, reviewable JSON artifact.

Source

The model does not read the transaction rows or decide that a discrepancy is probably fine.

Source

With inference disabled, the same mapping checks 52 generated weeks: 41 clean, 11 held for review, 19 explicit exceptions. Zero model calls during replay.

Source