A phone call becomes a structured spreadsheet row — in about 16 seconds, with zero human touch 一通电话,约 16 秒后自动变成一行结构化表格数据——全程零人工
Production system for donation-call intake: every incoming call is recorded, transcribed, understood by an LLM, and written to Google Sheets before the caller has put their phone down. 捐赠来电登记的生产系统:每通来电被录音、转写、由 LLM 理解并写入 Google Sheets——来电者刚放下电话,数据已经入表。
The problem问题
The client received donation pledges by phone. Every call meant a human listening, writing down the amount, the processing fee, and the beneficiary charity — slow, error-prone, and impossible to scale past business hours. 客户通过电话接收捐赠承诺。每通电话都需要一个人来听、记下金额、手续费和受益机构——慢、易错,而且下班之后就完全没人管了。
The constraint约束
No staffed call center, no CRM, no appetite for new software. The output had to land somewhere the team already lived: a Google Sheet. And the system had to run unattended — self-activating on every incoming call, day or night. 没有客服团队、没有 CRM、也不想学任何新软件。产出必须落在团队本来就在用的地方:一张 Google 表格。而且系统必须无人值守——不分昼夜,每通来电自动触发。
The solution方案
A self-contained daemon on a Linux server, wired into Twilio's webhook: Linux 服务器上一个自足的守护进程,接入 Twilio 的 Webhook:
- Call in → Twilio records the voice message and fires a webhook to the server. 来电 → Twilio 录下语音留言,并向服务器发出 Webhook。
- Transcribe → the recording is downloaded and converted to text by an ASR engine. 转写 → 录音被下载,并由 ASR 引擎转成文字。
- Extract → an LLM pulls out the structured fields: donation amount, processing fee, beneficiary organization. 提取 → LLM 抽取结构化字段:捐赠金额、手续费、受益机构。
- Deliver → a complete 8-column row (timestamp, caller, recording link, transcript, extracted fields, confidence) lands in Google Sheets. 入表 → 完整的 8 列数据(时间戳、来电号码、录音链接、逐字稿、提取字段、置信度)写入 Google Sheets。
End-to-end time from hang-up to spreadsheet row: ~16 seconds. 从挂断电话到数据入表,端到端耗时:约 16 秒。
The part that makes it trustworthy: confidence scoring让它值得信赖的关键:置信度评分
LLM extraction is never blindly trusted. Every row carries a 0–100 confidence score computed from explicit rules — deductions for an unstated amount, an unclear charity name, hesitation in the caller's voice, or poor audio clarity: LLM 的提取结果从不被盲目采信。每行数据都带一个由明确规则算出的 0-100 置信度分——金额未明说、机构名称含糊、来电者语气犹豫、音质差,都会扣分:
| Score分数 | What happens处理方式 |
|---|---|
| ≥ 90 | Auto-committed, no review needed自动入库,无需复核 |
| 70–89 | Human spot-check人工抽查 |
| < 70 | Flagged for manual review标记,转人工复核 |
The AI does the work; the score decides when a human needs to look. That's the difference between a demo and a system you can run a business on. AI 负责干活,分数决定什么时候需要人看一眼。这就是"演示品"和"能拿来跑生意的系统"之间的差别。
Proof证明
Shipped and running in production. Delivered with a handover document covering: how to start and monitor the daemon, the exact Twilio console settings (including the media-auth setting that silently blocks recording downloads), the full field reference, and a step-by-step guide for migrating every credential to the client's own accounts. The client owns the system outright — no dependency on me. 已交付,运行于生产环境。随附完整交接文档:守护进程如何启动与监控、Twilio 控制台的精确配置(包括那个会静默阻断录音下载的媒体鉴权开关)、完整字段参考,以及把每一个凭据迁移到客户自有账号的分步指南。系统完全归客户所有——不依赖我。
Where else this applies还能用在哪
The same pipeline works anywhere a phone call needs to become structured data: missed-call capture for contractors, after-hours intake for service businesses, order lines, appointment requests. If your business runs on phone calls that someone has to write down — this removes the writing down. 同一条管道适用于任何"电话需要变成结构化数据"的场景:承包商的漏接来电捕捉、服务型企业的下班后登记、订单热线、预约请求。如果你的生意依赖"有人一边听电话一边做记录"——这套系统把"做记录"这一步整个删掉了。
← All case studies← 全部案例