"Turpentine VC" | Venture Capital and Investing
00:28:14 · Usman Hanif, Brendan Foody
[00:28:14] S1: So if you're if you're offering AI, you're also citing your sources to almost ensure that the AI is working. [00:28:20] S2: Exactly, exactly. It's, it's really, really important, right? Like, think customers over time can build a lot of trust in us. But it, it makes sense that they would want to double check, especially as they're getting started. And so we really facilitate that versus other players in the space, which we talked about earlier, you know, have so much branding as AI companies. And we've talked to, you know, executives at all of these companies that corroborate that they're not AI companies, they just talk about it. And they're not actually automating these manual processes because they've built their entire businesses around services, not around product. And so I think keeping this, like, maniacal focus on complete automation and showing to customers the automation that we're doing is really important. [00:29:08] S1: Yeah. And you mentioned sort of the international talent space almost being a niche for you guys. Would you say that is sort of the wedge right now? What are most people using you guys for? Is it to hire folks internationally? And then if so, what type of roles are they hiring?
00:29:08 · Usman Hanif, Brendan Foody
[00:29:08] S1: Yeah. And you mentioned sort of the international talent space almost being a niche for you guys. Would you say that is sort of the wedge right now? What are most people using you guys for? Is it to hire folks internationally? And then if so, what type of roles are they hiring? [00:29:26] S2: Yeah. So we hire. We initially hired a lot of people, particularly in India, software engineers in India, then we expanded to all countries. And now we're expanding to all domains. So while majority of our business is still software engineers, and we've hired well over 1,000 software engineers, It's now expanding into many, many other domains, all sorts of business roles and to medical legal roles, etc. And the reason is that the technology is really broadly applicable. When you are able to conduct a really good interview of a candidate, and you're able to evaluate the resume really well, it's it's applicable to a lot more than just soft registering. And actually, the other thing I'll mention on this, which is like a really interesting technical point, is that there's a lot of transfer learning associated with making predictions across different kinds of models. And that if we make a model really good at predicting performance reviews for software engineers, it's actually much easier for it to learn how to predict who's going to be a good lawyer or a good doctor with far less data. And so this really goes to show how interesting the usage data network effect is, as we expand all these new roles, all these new disciplines within the economy.
00:30:40 · Usman Hanif, Brendan Foody
[00:30:40] S1: Yeah. That was actually gonna be my my next question about sort of how easy it is to move across domains. Right? If you're fine tuning these models, do you have to keep it specific to, say, software engineers? And I'd also be curious for you to touch upon that a little bit. I mean, it sounds like if you're taking into account data points, like, how nice... Like, how well their camera's working, it maybe speaks to intangibles more than even something specific. How how do you see that sort of playing out for all industries? Like, how do you expand to hiring lawyers on top of software engineers? [00:31:13] S2: Yeah. So we have to create labels and new domains such that we're, like, well positioned for them. And it's difficult to, like, say we wanted to add a new entirely unique domain without any labels, the models would probably struggle a little bit. However, if there's some overlap in domains, or we add a small amount of, like fine tuning data to the models at different parts of the stock, it's able to unlock a lot of new abilities. And so that can be really, really exciting. And yeah, it's super contextual to the So there's, like, scaling laws around all of this kind of stuff. And the more data that we get on software engineers, the better the model gets at evaluating all other roles. [00:32:03] S1: So serving a new role would sort of require collecting a bit more data on land and perhaps fine tuning again.
00:32:03 · Usman Hanif, Brendan Foody
[00:32:03] S1: So serving a new role would sort of require collecting a bit more data on land and perhaps fine tuning again. [00:32:11] S2: But oftentimes, not not as much as people would think. It... It's hard for me to give a really precise answer because it's so dependent on the role. Right? Like, as an example, say you have a lot of data on what makes a good full stack engineer. And now you want to predict what makes a good backend engineer, there's, you know, not too much nuance that it's like pretty easy for the transfer learning to apply to such a similar role. Versus if you want to predict what makes a good designer, it's going to really struggle. And so it's very dependent on like how related the attributes that you'd look for are in a given role to the attributes that you would look for in that new role, with some roles being quite challenging and others being much easier. [00:32:59] S1: We've touched upon this a bit when I asked how our company sort of vetting your vetting, like, making sure that you're doing it right. But you mentioned, like, you need data on what makes a good designer. Right? Say you've trained the model, you get an output. How do you just internally validate like, hey, this model is doing well or not doing well. This is actually a good designer. This is actually not a good designer. Is it pretty manual? Like, do you yourself just need to kind of do it by inspection? Or how do you do this systematically?
00:32:59 · Usman Hanif, Brendan Foody
[00:32:59] S1: We've touched upon this a bit when I asked how our company sort of vetting your vetting, like, making sure that you're doing it right. But you mentioned, like, you need data on what makes a good designer. Right? Say you've trained the model, you get an output. How do you just internally validate like, hey, this model is doing well or not doing well. This is actually a good designer. This is actually not a good designer. Is it pretty manual? Like, do you yourself just need to kind of do it by inspection? Or how do you do this systematically? [00:33:29] S2: Yeah, so we do occasionally have, you know, human teams that will use the product and find like, if there is a result that shouldn't be there, why it's getting surfaced, how can we fix that? And then I think the most important one is really the customer data of customers, you know, telling us giving us preference data of who they want to hire for which reasons of whether that's that person is performing well. And as the customer data begins to scale, we're leaning less and less on our own human teams that can help create data. [00:34:01] S1: Yeah. We've talked a whole lot about how, at least from the side of the employers, this is amazing. Let's, see a lot more candidates, you know, sort of offloads the vetting process. You also briefly touched on from the side of the candidates. Could you speak a bit more about how this sort of helps them? I'm sure a lot of folks, especially overseas, are actually really loving something like this that's legitimate, systematic, and works.
00:34:27 · Brendan Foody
[00:34:27] S2: Absolutely. I mean, there are so many people in The US, let alone abroad, that have applied to dozens, if not hundreds of jobs, and just not heard back from the vast majority of them because of extremely irrational resume screens. They don't get feedback on why they're not getting considered for opportunities. And the process is wildly unfair. But when you're able to lower the cost of assessing talent to zero, and do so in a centralized fashion, where they apply to one platform that can facilitate transactions with hundreds of companies, it just makes it a way better experience. So I think the two things client that are three things that candidates love about our platform are first of all, we consider everyone in a fair way, not just based on the credentials on their interviews, or on their resumes. The second thing is that candidates are able to get feedback on their resumes, on their interviews, why they're performing well, and selected for a role. And then the third thing is that they apply to one place that considers their applications for hundreds of companies that are hiring for 1000s of roles. And so it's just a far, far more efficient structure than this, like, very disaggregated labor market that we see today.