跳到正文
DEV Community· Hirescoot·· 2 小時前AI 評分38

Hirescoot 創辦人打造代求職者申請職位的 AI 智能體:表格比 AI 更難應付

I built AI agents that apply to jobs. The forms were harder than the AI.

AI 導讀

Hirescoot 創辦人打造 AI 智能體,協助求職者在僱主官網搜尋職位、撰寫履歷及提交申請,並以僱主確認電郵作為申請完成憑證。系統每日讀取逾 100,000 個職位,約 4 分鐘撰寫一份履歷、約 2 分鐘填表及提交,目前只適用於美國職位。每份申請均有屏幕錄影,只有申請者可以觀看。

正文

I am the founder of Hirescoot, so read this as a builder's account of what I ran into, not as a neutral review.

The job system is broken. People are being laid off because of AI, and employers use AI to filter out the rest of us. In the last semester of my master's I sent hundreds of applications, and nothing came back.

So I tried the AI job appliers, several of them. All of them fell short. They pull their jobs from third party boards, so my money went on applications to jobs that were probably not there anymore, and none of them showed real proof that it had applied. The tailored resumes were full of slop, with wrong information about me. Automated form fillers helped, but they are too rigid to use hundreds of times: they always stop at something they cannot handle.

So I decided to build my own: AI agents that find jobs on employers' official websites, write a resume for each one from your real facts, sign up on the employer's site for you, fill in the form and send it. You get the employer's own confirmation email. I expected the AI to be the hard part. It was not. The forms were.

Here is one application from start to finish, in 29 seconds:

Here is what watching software apply to jobs taught me.

1. Autofill stops at the first box it does not know

A profile has your name, your email and your last job. A real form asks "Why do you want to work here?", "Are you willing to relocate?", and then the same work history again in its own boxes. Autofill fills what it recognizes and leaves the rest blank or wrong. To finish a form, something has to read each question and work out the answer from what it knows about you, the way you would.

2. Every employer wants its own account

Before the form there is a sign up page, a password rule and often an email to confirm. Fifty applications can mean fifty accounts. This is where most people give up for the evening. The agent makes the account, and the password goes into an encrypted vault that belongs to the user.

3. One resume for every job is the real spam

Auto apply tools have a bad name, and they earned it: the same resume sent to hundreds of postings helps nobody. So each job gets its own resume, written from the person's own facts. Nothing is made up. Then a second agent scores the draft and sends weak ones back to be written again. Only a draft that passes goes out.

4. "Sent" means nothing until the employer says so

A form can fail quietly: a required box, a timeout, an upload that did not take. I stopped counting an application when the agent pressed the button. It counts when the employer's own site confirms it, and the employer's confirmation email goes to the person who applied.

5. People do not trust what they cannot see

This is the one I did not plan for. If software applies in your name, you want to know exactly what it wrote. So every application has a screen recording. You can watch the form being filled in on the company's own website, answer by answer. Only the person it was sent for can watch it.

What I decided not to build

Three things were easy to build, and I left them out.

A button that sends one resume to every posting. It is the fastest way to a big number and the surest way to no replies.

Made up experience. The writing agent works only from what the person told it. If a job asks for something they have not done, the resume does not claim it.

A black box. If I cannot show a person what was sent, I should not send it.

Where it stands

Today it reads more than 100,000 open jobs a day on employers' own hiring sites. It takes about 4 minutes to write a resume for a job and about 2 minutes to fill in and send a form. It works for jobs in the United States only, for now.

What I still think about

Employers use AI to read applications, and applicants now use AI to write them. I do not think more volume is the answer. I think fewer, better applications are, with proof of what was sent. Hirescoot opens in December, and there is an early list if you want to try it.

So, a question for people here who have hired or job hunted this year: if an agent applied for you, what would you need to see before you trusted it?

來源:DEV Community · dev.to