How we screened 1000s of resumes with this AI Automation workflow

Name: How we screened 1000s of resumes with this AI Automation workflow
Uploaded: 2025-06-15T15:29:41Z
Duration: 7 min 4 s
Description: Tired of spending countless hours manually reviewing resumes? In this video, we break down the exact AI automation workflow we used to efficiently screen thousands of job applications at SkilAgents.ai.

Video 2 of 2 · 7:04

Transcript

Auto generated by YouTube. Click any timestamp to jump to that moment.

Show

0:03in the last 3 months. Um we had
0:06plus internship profiles coming
0:09right and center. Um unfortunately
0:13fortunately because we realized there
0:15a problem. We hosted all this via
0:17and LinkedIn has a very
0:19UI. You cannot download all
0:20rums. They give you a very
0:22screen where you browse each of
0:24element each of those profiles one
0:26one. It was way too time consuming.
0:29now, what have we done here? Um,
0:33show you my Google Drive
0:44one. It'll extract all the data from
0:47PDF. your email, where did you work,
0:50college are you from, what was
0:51degree, where was your last work
0:53which was your last project,
0:55are your major skill sets, whatever
0:57mentioned in your um resumeum, right?
0:59extract all the data into very
1:01structured format. Then we run some
1:03like we have some requirements. We
1:05to hire engineers who are
1:07with some front- end backend
1:09as well as AI applications.
1:11on this criteria, it will score
1:12screen relevant candidates so that
1:15can go jump into interviews directly.
1:17I'll just show you uh how how this
1:19works, right? We'll start the
1:20So I have connected it to my
1:22drive. It has taken all my rums
1:25Google Drive. Now it's trying to
1:27all the data from my resume. I'm
1:294.1 model for this. The latest 4.1
1:32Chat GBD. It's extracting all the
1:35from the rumé. Then it goes
1:38something called as
1:39Uh it summarizes each of
1:41profiles, looks at all the data and
1:43comes up with a summary of what
1:44each profile means. and post
1:46It goes through each uh profile
1:49by one and starts dumping them into
1:51Excel sheet. So maybe we'll have a
1:53of them. So the first one got
1:54We got the name, mobile number,
1:57email, college, degree, major,
2:00work experience, project. What are
2:02skill sets? What are summary? What
2:04the summary of his profile? How is
2:05compatibility with respect to the JD
2:07I floated? Probably very less
2:09the candidate hasn't done much
2:10AI applications. Uh but he has done
2:12in front and back end but not
2:13on AI. So his rating is pretty low.
2:15it'll go through all the 10
2:17that I have. It'll start making
2:20Again, this guy is from a
2:22college, computer science
2:23This guy seems to be rating
2:25And why is he rated high? Mostly
2:27he has got strong technical
2:29and he has proven AIM project
2:32and proficiency, right? He has
2:34done something on lang chain MCP or
2:36which I wanted. That's why he's not
2:38out of 10. But otherwise this guy
2:40to have done something couple of
2:41on LLMs in his projects. We'll
2:44a lot of these candidates. I mean
2:46of those profiles will be passed
2:48dumped into this Excel sheet in this
2:49format. What we can also do
2:51based on the relevancy score I can
2:53send rejection emails to
2:54of the candidates. Send a follow
2:56calendar invite to candidates
2:58are rating very very high in this
3:00score and keep on iterating.
3:02mean reduce my time in terms of
3:04and directly go and talk to
3:06of these guys, interview some of
3:07guys. What I'll show you quickly
3:09similar prompt structure in terms of
3:12we used. As I told you, there is a
3:14prompt, there is a system prompt in
3:16particular example. If you see, we
3:18passing the job title, we are
3:20the summary in terms of what we
3:22before. Key responsibility is
3:24come this comes from my JD and we are
3:26it to actually score the
3:28right? How does he rate? How
3:30you score this candidate is what
3:32trying to do here. I'll maybe show
3:34more system prompt in terms of
3:36very simple prompts right
3:38concise summary based on all this data
3:41the AI was able to come back with
3:43profile summary that we saw there
3:45these are very very basic prompts
3:47we are using you guys can actually
3:48this to a much much much
3:50version but this is good um out
3:52the box already probably I would rate
3:54at 85 to 90% accuracy let us quickly
3:56at um this still running because I
3:5810 documents I have added a wait
4:00very deliberately because Google
4:02has a limitation rate limiting
4:04it doesn't permit when I run
4:06API altogether but again Google
4:08is only for demo you can put this
4:10in a database for a more structured
4:12and then passing it for emails
4:14etc so so for enterprise
4:16software probably workflows
4:18be slightly different since it's a
4:20we are using Google sheet to
4:21art of what is possible looks
4:23the extraction is complete we got
4:25of files out here 10 odd files I'll
4:28one of the examples yeah the resume
4:31passed I'll take one of the examples
4:33probably go through what the passing
4:35was looks Looks like this guy
4:37got lot of data. Let's go check his
4:40Okay, this is Ani Singh start
4:42number starting with 80 ending
4:4489.
4:46Singh mobile 80 and 89 email
4:52communication 2025 batch
4:562025 batch. And maybe let's look at
4:59as well. This guy has got
5:01places where he has worked. He
5:04worked at Ragam as a tech head which
5:06to be his latest workx. So let's
5:08what it picked. Yeah, it says Ragam
5:11head from it also passed a date.
5:13from 1st September 2024 to 1st Feb
5:17The sky is the limit guys. So I
5:19actually passed only the first
5:21uh the latest exper work
5:23and some data. So as you can
5:25it got all all the data accurately.
5:28got skill sets etc. All these are
5:31perfectly fine. Very
5:33points here is we're not
5:34any keyword matching like typical
5:36systems work. What we're doing is we
5:39passing all this context to LLM as
5:41individual person would do. The LLM
5:44um acting like a person here, right?
5:46reading all this data and then it's
5:48to its own conclusions based on
5:50criteria I have set. If you look at
5:52experience, the order in which he
5:55is not I mean it's going in the
5:57chronological order. He was
5:59from the farthest first and
6:00at the end. But I asked LLM to
6:03give me only the most
6:04experience and hence it p pulled
6:06and not tatwa over here. Similarly,
6:09can see the mobile number has
6:11code and um typically excel
6:13if you add country code etc. I
6:15LLM to ignore some of these
6:17codes so that mobile number can
6:19passed properly. It's able to do that
6:21simple instruction. Right? I did
6:22add any reax or complicated logics.
6:25just a system prompt. It made all
6:27possible. What I can also do given
6:29has given his GitHub URL, LinkedIn
6:31and all those. I can actually make
6:34agents go to his GitHub, check
6:36he has built and then rate his
6:38and then come back and tell me
6:40skill sets based on whatever his
6:41repositories are or go to his
6:43read his post, understand what
6:45really curious, interested about.
6:48you guys. Um that's the end of
6:50session. Next week we'll be
6:53uh model context protocol and
6:55dashboards. See you in the next
6:57Bye-bye.

Claude Just Did My Entire Sales Job in 8 Minutes | This Actually Happened

Want the next one in your inbox?

Join 1,000+ Product Managers getting one deep dive every Friday.