# Autocomplete for Actions: System One from the UI to the DB

Source: https://www.youtube.com/watch?v=ZZ19zhYnDKY

Machine-generated transcript; recognition errors remain. Source-recording timestamps. Boundary-crossing subtitle ends are clipped to this session.

[06:24:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23082s) All

[06:24:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23087s) right. Hey everyone.

[06:24:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23090s) I promise I'm the last uh person on the

[06:24:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23092s) hype train for Jev. Uh I'm Scott. I'm

[06:24:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23096s) from uh I'm from Workday and I'm going

[06:24:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23098s) to go through how we're experimenting

[06:25:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23100s) with Jev uh to drive uh new innovations

[06:25:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23104s) in HR and finance tasks. Uh exciting

[06:25:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23108s) stuff. So, uh, at Workday, we have over

[06:25:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23112s) 11,000 different organizations that use

[06:25:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23114s) our tools. And a lot of the problems

[06:25:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23116s) that we deal with break down into kind

[06:25:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23118s) of two different categories. So, one is

[06:25:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23121s) very quick actions they have to take. I

[06:25:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23123s) want to see my pay. I want to ask for a

[06:25:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23125s) day off. The other ones are I have to

[06:25:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23128s) dive into a ton of data, weed through

[06:25:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23130s) it, try to figure out what's going on.

[06:25:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23133s) And like many of your companies as well,

[06:25:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23134s) we're trying to, you know, put agentic

[06:25:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23136s) workflows around this. And so if you

[06:25:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23139s) look at just a a very simple uh agentic

[06:25:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23142s) loop, you know, we have a user request.

[06:25:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23144s) We try to decide what to do. We're

[06:25:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23146s) running tools. We're judging did we do a

[06:25:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23148s) good job? Uh and then over and over

[06:25:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23150s) again until we finally get a result. Uh

[06:25:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23153s) now one of the problems here is that on

[06:25:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23155s) each one of those we're doing LLM calls

[06:25:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23158s) and there's two major problems with

[06:26:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23160s) that. So one is there's too much

[06:26:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23162s) intelligence. So if we look at the

[06:26:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23164s) system one versus system two thinking uh

[06:26:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23168s) we're doing a lot of slow deliberate

[06:26:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23170s) very expensive thought processes at each

[06:26:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23173s) one of those steps.

[06:26:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23176s) Another problem is it's just the wrong

[06:26:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23178s) tool for the job. So you know LLMs a lot

[06:26:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23182s) of times are just a hammer uh looking

[06:26:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23184s) for a nail. And you know we want to

[06:26:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23186s) figure out like what do we actually

[06:26:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23188s) need? Uh because clearly LLMs are not

[06:26:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23191s) doing it. So we want something that is

[06:26:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23194s) fast

[06:26:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23195s) uh smart easy to use uh it's generalized

[06:26:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23200s) and it's very very cheap.

[06:26:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23204s) So when we look at like from a code

[06:26:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23206s) perspective you know what are we really

[06:26:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23208s) looking at from a from a jevlike model

[06:26:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23210s) it's really just a conditional if this

[06:26:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23213s) happens this should happen. Which one of

[06:26:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23216s) these cases are the ones that I want to

[06:26:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23218s) explore?

[06:27:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23220s) And you know you've seen it among many

[06:27:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23223s) of these presentations already. Uh but

[06:27:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23225s) Jev has the primitives to allow us to do

[06:27:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23227s) this. Uh there's the probability is it

[06:27:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23229s) you know this or that and what's the

[06:27:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23231s) probability? Uh what are the different

[06:27:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23233s) categories that this uh particular thing

[06:27:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23235s) could fall into? And then you have the

[06:27:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23238s) uh the scoring mechanism which is

[06:27:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23240s) basically probability and sorted.

[06:27:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23244s) So you know this has been brought up as

[06:27:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23247s) well. You know, my first thought when I

[06:27:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23249s) saw, you know, the Jev models was, well,

[06:27:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23251s) we already have classifiers. I've been

[06:27:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23253s) doing this for a long time. How is it

[06:27:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23255s) actually different?

[06:27:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23257s) And as many of the speakers have talked

[06:27:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23259s) about already, you know, with

[06:27:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23260s) classifiers, I need to figure out, you

[06:27:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23263s) know, what's my golden data set. I have

[06:27:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23266s) to train it. New data comes out. I have

[06:27:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23268s) to iterate over and over. With Jev, just

[06:27:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23271s) out of the box, it just worked. We

[06:27:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23273s) plugged in our question. We classified

[06:27:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23276s) it. we didn't really have to do

[06:27:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23277s) anything. It generalized to our

[06:27:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23278s) question.

[06:28:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23280s) Uh and so we started using it for a lot

[06:28:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23282s) of the use cases that you've seen from

[06:28:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23284s) some of the presenters already. Uh so we

[06:28:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23286s) use it for routing. That's kind of the

[06:28:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23287s) canonical example. Um you know figuring

[06:28:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23290s) out did this step work you know as a

[06:28:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23292s) judge. Uh we looked at uh you know

[06:28:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23296s) resumes you know is this person

[06:28:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23299s) qualified for ML work? Uh, and I want to

[06:28:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23303s) do kind of an overly simplified example

[06:28:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23305s) as well. Um, so no one taught Jev like

[06:28:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23308s) what a fish is. And so this is actually

[06:28:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23310s) a question we asked like is a fish a

[06:28:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23312s) pet? Well, 50/50 might be a pet

[06:28:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23315s) sometimes, might not be. Um, and so, you

[06:28:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23318s) know, if I ran into this, then I have to

[06:28:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23320s) figure out um, what do I do next? Maybe

[06:28:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23323s) I can give it some more context. Oh, is

[06:28:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23325s) a fishing a bowl a pet? You know, and

[06:28:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23327s) after 20 milliseconds, I can ask this

[06:28:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23329s) again. All right, that's great. Oh, is a

[06:28:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23332s) fish on my plate a pet? Clearly not. Uh,

[06:28:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23335s) so this is something that wasn't trained

[06:28:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23337s) out of the box in seconds. It just

[06:28:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23339s) worked.

[06:29:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23342s) All right, so now on to some of the

[06:29:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23343s) demos that we're do that um, you know,

[06:29:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23346s) we've done at workday. So just a caveat

[06:29:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23349s) on some of these. Um, it's not real

[06:29:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23352s) data, so we're not using any of your you

[06:29:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23354s) know, resumes or payroll data. Uh, these

[06:29:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23356s) are also not in production now and may

[06:29:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23358s) or may not be in the future. So that's

[06:29:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23360s) something our legal asked me to say. So

[06:29:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23362s) let's get into, you know, some of the uh

[06:29:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23365s) the demos.

[06:29:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23367s) All right. So starting with

[06:29:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23368s) autocomplete. So autocomplete starts to

[06:29:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23371s) predict what you're trying to ask. It's

[06:29:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23374s) just predicting the question. It's not

[06:29:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23375s) helping you get to the answer. So as you

[06:29:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23378s) start to type more, it's going to guess

[06:29:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23380s) like, oh, I want to figure out if this

[06:29:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23383s) person wants to take a day off work.

[06:29:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23384s) Great. You've guessed the question, but

[06:29:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23386s) you really haven't, you know, helped me

[06:29:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23388s) get to the answer.

[06:29:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23390s) So what we want to do is move from just

[06:29:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23392s) predicting the word or the next block of

[06:29:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23394s) code and actually help you finish the

[06:29:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23396s) task. And so this is best shown with a a

[06:30:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23400s) demo that we have where we go from

[06:30:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23402s) predicting what you want to say to

[06:30:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23405s) predicting what you want to do. So

[06:30:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23408s) hopefully the video play. There we go.

[06:30:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23410s) All right. So I need to take time off of

[06:30:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23413s) off of uh next week. And as I start to

[06:30:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23417s) ask this, Jev is actually figuring out

[06:30:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23419s) under the hood, okay, what what more

[06:30:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23422s) context do I have? How can I adjust this

[06:30:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23425s) UI based on the information that I'm

[06:30:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23428s) that I'm getting? This is not something

[06:30:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23430s) that we could have done with an LLM.

[06:30:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23431s) We're getting, you know, we're we're

[06:30:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23433s) doing this, you know, in a loop while

[06:30:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23435s) someone's asking. Um the pay slip, you

[06:30:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23439s) know, you shouldn't have to go in and

[06:30:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23440s) ask an agent, hey, can you go find out

[06:30:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23442s) how much I was paid? Uh what you can do

[06:30:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23445s) is just in the search say what was my

[06:30:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23447s) pay last period

[06:30:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23450s) pops up an interface allows you to

[06:30:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23453s) unhide it you know see what your pay was

[06:30:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23455s) in the last period very simple gets you

[06:30:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23459s) to the action book a meeting.

[06:31:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23462s) So I know that I want to book a meeting

[06:31:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23465s) and uh I'm going to want to do it for

[06:31:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23468s) next Friday. Let me do it at 2 PM. Oh, I

[06:31:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23472s) want to do it with my boss. Make sure

[06:31:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23474s) that I get my, you know, my boss on

[06:31:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23476s) this. Uh, Logan McNeel. There we go.

[06:31:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23479s) Meeting's booked.

[06:31:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23481s) And then a final one that we have just

[06:31:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23483s) for this example, but there's many more.

[06:31:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23486s) Uh, is finding a person. So, in this

[06:31:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23489s) one, I may have some semantic knowledge

[06:31:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23491s) like find the person who owns this part

[06:31:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23494s) of the organization or just like who's

[06:31:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23497s) Ashanti? I want to contact her. uh pop

[06:31:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23499s) it up. Click on email. Good to go.

[06:31:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23506s) All right. So, what's happening under

[06:31:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23507s) the hood? So, because Jev is fast and

[06:31:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23510s) cheap, uh just like a type ahead, we can

[06:31:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23514s) start to render a UI on this on that

[06:31:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23516s) same cadence, we can look at as the

[06:31:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23519s) words are coming, do a prediction of

[06:32:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23521s) intent and start to render an interface.

[06:32:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23525s) And you know we could do this at you

[06:32:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23527s) know about 20 bucks for a million

[06:32:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23528s) judgments. Uh and the way that we're

[06:32:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23531s) doing it is kind of twofold um or two

[06:32:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23534s) parts I should say. So first Jev is

[06:32:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23537s) picking the card of saying what's the

[06:32:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23538s) intent. All right there's a card that's

[06:32:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23540s) going to help you out and then the team

[06:32:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23542s) that's responsible in that domain then

[06:32:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23544s) can decide what are the questions to

[06:32:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23546s) progressively fill out that card. And so

[06:32:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23549s) there's sort of this two-step process.

[06:32:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23551s) Uh the first step is getting people in

[06:32:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23553s) the front door, getting the right card

[06:32:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23555s) up, and then the second one is making

[06:32:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23557s) sure that you fill in that card to help

[06:32:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23559s) them accomplish the task that they're

[06:32:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23561s) after. Now, we have an eval on here

[06:32:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23564s) before we merge any card. We want to

[06:32:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23566s) make sure that it doesn't collide with

[06:32:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23567s) any other one. And at least in our

[06:32:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23570s) initial experiments, you know, we we've

[06:32:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23572s) done seven cards or a just chat because

[06:32:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23575s) it should if there's no card to actually

[06:32:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23577s) solve it, it should allow you to just

[06:32:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23578s) continue chatting. uh we were able to

[06:33:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23580s) get you know 96% accuracy on that.

[06:33:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23585s) All right. So now on to the recruiter.

[06:33:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23587s) So let's talk about the recruiter

[06:33:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23588s) problem. All right. So this is not like

[06:33:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23590s) a fast uh task. It's something where

[06:33:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23593s) recruiter has a pile of hundreds of

[06:33:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23595s) resumes and has to figure out who has

[06:33:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23598s) the best resume to accomplish the uh or

[06:33:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23601s) fill the wreck that they're trying to

[06:33:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23602s) fill. Now I first want to look at like

[06:33:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23605s) how we would have done this before. uh

[06:33:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23607s) we would have done you know an embedding

[06:33:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23609s) put the vectors in a vector DB then had

[06:33:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23613s) a retrieval layer then have a reranker

[06:33:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23615s) and then figure out based on the

[06:33:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23617s) reranker which ones that we want to put

[06:33:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23619s) in front of uh the recruiter. Uh one

[06:33:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23622s) interesting call out is the Quinn

[06:33:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23624s) reranker uh looks surprisingly similar

[06:33:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23627s) to uh the uh the Jev process as well. Uh

[06:33:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23632s) and so that that's kind of an

[06:33:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23633s) interesting correlary if you've used

[06:33:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23634s) that one before. Uh but with Jev, we

[06:33:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23638s) basically take out all the embedding

[06:33:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23639s) steps. I don't have to train anything

[06:34:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23642s) up. I don't have to do embedding. I

[06:34:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23643s) don't have to put things in a vector

[06:34:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23645s) database. Um what I can do is just do a

[06:34:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23650s) function on my database. So put all the

[06:34:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23653s) resumes into a database. Uh use natural

[06:34:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23657s) language to query it, iterate over every

[06:34:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23660s) single one of them, and then just figure

[06:34:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23662s) out which ones match my criteria.

[06:34:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23665s) So uh I'll go into an uh you know an

[06:34:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23668s) actual demo of this exact same one. So

[06:34:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23671s) here we have a select statement and so

[06:34:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23673s) this is against a duct db where we have

[06:34:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23676s) a function that is going to call jev to

[06:34:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23680s) figure out what's the probability that

[06:34:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23681s) someone matches this criteria. We're

[06:34:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23683s) going to get it a score and then we're

[06:34:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23685s) going to order by that score. Now I'm

[06:34:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23686s) going to do about 20 and as I said these

[06:34:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23689s) are all fakes and none of your resumes

[06:34:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23690s) are in here I promise. And then here we

[06:34:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23693s) see uh these are very uh ML focused uh

[06:34:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23696s) synthetic data as well. So it's not a

[06:34:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23699s) surprise that we would have uh very high

[06:35:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23701s) hit rates against someone who's worked

[06:35:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23703s) in fraud, trust, safety problems. Um but

[06:35:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23706s) let's look at let's change this

[06:35:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23707s) question. Let's uh be a little bit more

[06:35:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23709s) adversarial on this one. So a lot of

[06:35:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23712s) these résumés have the word pipeline in

[06:35:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23714s) them. So whether it's a data pipeline or

[06:35:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23717s) an AI pipeline, uh pipe's going to come

[06:35:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23720s) out. So let's just ask like is this

[06:35:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23722s) person qualified to do plumbing? If you

[06:35:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23723s) were going to just do a lexical search,

[06:35:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23725s) it might pick up some of these. So we

[06:35:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23728s) put it into Jev and asked, you know, of

[06:35:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23729s) all these ML res resume résumés, let's,

[06:35:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23732s) you know, trim the limit down to

[06:35:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23734s) five here. Uh go ahead and run it.

[06:35:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23738s) And we can see, you know, that um very

[06:35:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23743s) very low probability on these. So these

[06:35:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23745s) are ones that we would almost definitely

[06:35:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23747s) throw out. And you can see the speed. So

[06:35:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23749s) we went over 117 records. Uh we had 2.3

[06:35:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23753s) seconds of wall time and spend almost

[06:35:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23755s) nothing on this query to just iterate

[06:35:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23757s) over everything. Uh so Jev really kind

[06:36:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23760s) of changed the game for us because every

[06:36:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23762s) every place where we're thinking either

[06:36:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23764s) either this is too expensive or it's too

[06:36:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23766s) slow, we thought of new places to inject

[06:36:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23769s) Jev into our pipeline.

[06:36:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23771s) And so [snorts] just to summarize, you

[06:36:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23772s) know, the two areas that I went through

[06:36:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23775s) are one, where can we decide very

[06:36:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23778s) quickly and get to the intent and get

[06:36:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23780s) the person on their way? And then the

[06:36:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23783s) second one is to cut down on some of

[06:36:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23785s) the, you know, really deep work. How do

[06:36:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23788s) we filter those out for, you know, our

[06:36:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23790s) customers? And so I hope this that that

[06:36:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23793s) this gave you a little bit of insight

[06:36:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23794s) into the things that you can take back

[06:36:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23796s) to you know your own companies and how

[06:36:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23798s) and you know imagine the things that you

[06:36:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23800s) can do with Jev and decision models as

[06:36:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23802s) well. And that's my time. Thank you very

[06:36:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=23804s) much.