Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

114,000 Kazakhs got a letter from the tax office. A local Gemma answers most of it for you.

说明与用途

sloth‑ledger 使用本地 Gemma 4 模型自动分类 100/3/12 税务规则下的转账,减少人工检查至 9 位发送者并生成回复信。

问题

Kazakh banks report a personal card to the tax office when it receives transfers from 100 or more different people in each of three consecutive months, totaling more than 12 minimum wages, prompting 114,000 people to receive a notice asking them to explain the source of the money.

结果

On a sample statement of 537 operations and 205 senders, sloth‑ledger completed the categorization in 17 minutes on an RTX 5050 laptop, leaving only 9 senders for human review and producing a draft reply letter with accurate totals. The approach eliminates manual transfer review and reduces the risk of incorrect totals in the tax letter.

工作流

  • 1. Extract PDF statement using regex over extracted text.
  • 2. Count senders per month and verify 100/3/12 rule with plain Python.
  • 3. Group transfers into unique senders (≈205) using plain Python.
  • 4. For each sender, create a compact profile (transfer count, active months, median/range amounts, days of month) and pass it with owner notes to Gemma 4 running locally in Ollama.
  • 5. Receive JSON output (category, confidence, reason) from the model, enforce guard rails to drop invented senders, mark low confidence or malformed JSON as "unclear", and collect uncertain senders for human review.
  • 6. Generate a draft reply letter by filling a template with the verified totals and sender categories.
  • 7. Present the short list of uncertain senders and the draft letter to the user for final review and copy‑paste into the tax office portal.

证据

Since January, banks in Kazakhstan report a personal card to the tax office when it receives transfers from 100 or more different people in each of three months in a row, adding up to more than 12 minimum wages (1,020,000 ₸, about $2,000).

来源

sloth‑ledger does that sorting. You give it your card statement and two or three sentences in your own words: ... and it gives you back: the 100/3/12 check, month by month; every sender sorted into customer / family / shared expenses / loan / unclear, each with a one‑line reason; a short list of the people the model wasn’t sure about; a draft reply to the notice, where every number and every name comes from the statement.

来源

The model never sees the ledger and never computes the rule. It gets one compact profile per sender … and answers with a category, a confidence and a reason, constrained by a JSON schema.

来源

The column I care about most is the last one, and it’s zero for every model — because of the guard rails, not the models.

来源

On the sample statement (537 operations, 413 incoming transfers, 205 people): By hand: open every transfer, group by person, decide who each of 205 people is, add up totals per group, type the letter. With sloth‑ledger: about a minute to write three sentences, 17 minutes of the laptop doing it on its own, then a short list of 9 people to check.

来源

  • 记录日期: 2026-10-21T02:41:22.385Z