RateCon Radar: Gemma reads a trucker's contract in a browser tab, and the contract never leaves the laptop

说明与用途

RateCon Radar 在浏览器中本地运行 Gemma 4，解析运费确认书，标记潜在费用条款并提醒缺失条款，确保文档不离开用户设备。

问题

Freight carriers often sign rate confirmations under time pressure without fully understanding costly clauses, risking significant financial penalties.

结果

On a test set of eight synthetic rate confirmations, Gemma 4 identified 64 of 65 money clauses, correctly validated all 48 load facts, and produced 64 accurate plain‑English lines out of 77. The app runs locally in about 26 s per document on an Apple Silicon Mac, with no data leaving the device.

工作流

• Open the RateCon Radar demo in Chrome or Edge on a laptop or desktop.
• Click a sample button or drag a PDF into the page to load a rate confirmation.
• The app downloads the Gemma 4 model (3.1 GB) on first run, then loads it from local storage.
• pdf.js extracts text; the text is split into sentences and filtered for potential money clauses.
• Gemma reads each candidate sentence with 13 worked examples and outputs a label and a one‑line plain‑English explanation.
• A second prompt extracts load facts (broker, load number, pay, lane, dates, miles) as JSON and validates them against the document.
• The UI displays each flagged clause with the original sentence and the plain‑English line, and warns if the model’s interpretation conflicts with the quote.
• If the contract lacks detention, layover, or TONU terms, the app drafts a message asking the broker to add them.
• The user reviews the flags, edits the draft if needed, and signs the contract locally.
• The app includes a privacy test button that verifies no outbound requests are made.
• The entire process runs on the user’s GPU via WebGPU, with no server or backend.

证据

RateCon Radar reads the document before the driver signs it. You drop in the PDF. It pulls out the broker, load number, pay, lane, dates and miles, then lists every clause that can cost you money.

来源

The test set is eight made‑up rate confirmations rendered to PDF, with 89 labeled money clauses. Two documents were used while writing the prompt. The table below is for the other six, which hold 65 clauses and were scored once after the prompt was frozen.

来源

Gemma 4 E2B (3.1 GB) found 64 of 65 money clauses, 8 false alarms, 48 of 48 load facts right, 26 s per document.

来源