Notes from the build
What the data showed, what I got wrong, and what I changed because of it.
- We sent an AI agent to 100 SaaS websites. It found the price on 90, and how to cancel on 62.A browsing agent, three read-only questions, 100 well-known SaaS sites, every run recorded. Prices are easy to find. Cancel instructions are not, even when the official page exists.
- My PDF converter turned every borderless table into loose numbers. The fix runs on a CPU.I tested my PDF-to-Markdown tool against an ML converter on the same four files. It lost on tables. Here is what closed the gap without a GPU, and an n8n workflow for RAG.
- My app told its first merchant that $144 million was blocked. The whole catalog came to $383,000.Multiplying price by stock looked like the more honest number. Stock was the one input I could not check.