All Posts

How we built a batch audio transcription pipeline for 47,000 WhatsApp voice notes and videos — multilingual Southeast Asian speech, resume-safe, $19 total cost.

我买了一份五千多篇中文长文的数码化存档，然后用AI逐篇分析，提取出将近五千条关于作者本人的事实。碎片拼在一起，我看到了一个极其清醒、极其冷峻、但骨子里又极其有温度的人。

I had 21 scripts, 19 JSON files, and 80 percent coverage. After one afternoon of refactoring into a single CLI with SQLite and parallel Chrome instances, I hit 97 percent. Here is what I learned about turning script chaos into a real pipeline.

I used to teach top-down: big picture first, details later. Then I realized most people learn the other way around. This is a small shift in how I think about training.

I ran payroll with an AI agent. Not coding. Actual HR work. We caught 4 bugs, verified 10 payslips, and discovered that 40 percent meta-work overhead pays for itself in one cycle.