Original Summary

Chrome autofill stops at name, address and card. Everything past that you type by hand: the patient intake form at a new clinic, a school registration for your kid, a rental application, a phone number split across three boxes, a dropdown that wants "United States" on one site and "US" on another. I fill a lot of those, so in June I started building an extension that reads the whole page once and fills all of it in one click. It's called Fillaid. I want to be upfront that this took way longer than I expected, and most of the time went into measuring rather than building. What "measuring" meant I didn't trust "it looks like it works", so I built 33 test forms covering the common shapes (checkout, medical intake, tax, rental, job application, forms with a co-borrower or a parent section) and wrote an answer key by hand: 668 fields, each marked fill / leave alone / optional. Then I saved 15 real forms from live sites (city government forms, a Greenhouse job page, Jotform, a bank demo) and did the same for those. My first honest run on Aug 31 was bad. Precision 0.55. It filled 226 boxes it should have left alone: a loan amount box got a card number, "co-borrower email" got my email, honeypot fields got filled, and it happily answered every "Yes/No" question with Yes. That week was mostly writing guards: never touch a box in someone else's section, never write a value that isn't in the profile, leave checkboxes and radios alone unless it's obvious. The model part, and where I was wrong Obvious fields (Email, Phone) are matched by plain rules on your device. The messy ones go to a model. I fine-tuned a Qwen3 1.7B on 16k synthetic forms, served it on a rented GPU, and got it to 0.89 precision through the actual extension. I was pretty proud of that. I lost two days to a config bug where the merged weights silently picked up the wrong rope_theta and precision collapsed to 0.81 until I copied the base config back in. Then someone suggested I try Jev, TypeSafe's decision model. It's not a text model. You give it a state and typed questions and it returns one choice per question with a probability. So I wrote a small harness: for every field, one question "which of these saved values belongs here, or none". By construction it can't invent a value that isn't in your profile. Same forms, same answer key, same day, both models re-run. First number is my fine-tune, second is Jev: 33 test forms, precision: 0.888 vs 0.980 33 test forms, forms with every field right: 19 vs 25 15 real-site forms, precision: 0.870 vs 0.980 15 real-site forms, forms with every field right: 5 vs 12 Time per form: about 3.8 seconds vs about 0.8 seconds Cost per form: a GPU billing whether it fills or idles, vs $0.0006 Across both sets: 443 fields filled right, 9 wrong. My extension's own guards had been silently rejecting 202 of the fine-tune's raw picks; with Jev they only

  • 情报分类：工作与职业机会
  • 分类依据：内容涉及招聘、求职或职业发展
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/9/25 04:53:19