Invest with AI
00:24:56 · Sagar Kadakia
[00:24:56] S0: You could take it anytime you want to. That's the beauty of of the AI interviewing, right? I think um when you look, the biggest unlock here was ultimately removing all of the scheduling, right? Um because when you're like the the concept of scheduling itself will just limit the amount of calls that can be done, right? Overall, right? If we look at the universe of calls being done, right, it's only as fast as people can schedule and then take those calls. Well, what happens when you've removed those two massive bottlenecks and now people can participate whenever they want to? There's no time zone barriers, right? Someone can take it at 2 p.m., 6 a.m. It doesn't matter, right? I don't have to try to connect with someone, you know, uh uh uh in Singapore or in India or in London or in California. The AI interviewer is always ready. It's fluent in lots of languages. And so once you start peeling back all those barriers, right, it makes it a lot easier for folks to ultimately participate. They can participate when they want to. They can participate in the language that they want to, right? In the comfort of their own home on their iPad or their laptop or their or their mobile phone. And so that's how uh we're able to work at such a speed, right? Because think about it from this perspective. If you are a traditional network, all the scheduling has to be done from nine to five. Great. There's there's there's no concept of that. I feel like half the calls we do are before work, after work, Friday at 5 p.m. on the weekends, right? People submit projects on Friday, they're done on Monday. Right? Traditional networks can't do that. And so now we've we've we've massively expanded, right, the the amount of time uh that someone can ultimately participate. And so that's that's what that looks like.
00:26:35 · Brett Caughran
[00:26:35] S1: That's interesting. How do you protect against agent to agent? So uh I'm an expert, you reach out, your agent reaches out to me. I spin up my I drop in all my contacts into a voice agent and I have my agent log in and talk to your agent. Or is that a problem?
00:26:52 · Sagar Kadakia
[00:26:52] S0: Absolutely, it's a problem. Uh um, you know, uh fraud is a huge problem um in in in this industry, um, especially with the advent of AI tools, right? Um, someone could read off Chat GPT, sorry, someone could read off of Claude, right? And so those are real concerns. Someone can fake their expertise, right? So, how do you verify and authenticate them? I think one of the biggest pieces of feedback that we get from clients today is just how much fraud there is in the industry, right? How many times do you start talking to someone and you're like, I know you're reading off of something, or like you're definitely not an expert, right? You have some peripheral knowledge. And so that's actually AI interviewing is actually probably the best form factor for removing fraud. And I'll tell you why. Uh, when the AI interviewer is interviewing the expert, um, the way all the follow-ups work is obviously we record all the audio data that the expert is saying, right? Uh we transcribe it, it goes into an LLM. Well, guess what happens when it goes into the LLM? We're checking on whether or not that content was AI generated, right? And so if someone is talking, it's actually very easy for our interviewer to know uh whether or not that person is using uh tools because there's no inflections in their voice, it sounds like they're reading. Um, a lot of AI tools have very um uh they don't sound like humans. They do things like they list reasons out, right? Like humans don't talk like, you know, the reason I'm spending more is because our users are increasing, our workloads are going up, and XYZ is halfway. That's how ChatGPT talks. That's not how humans talk. And so uh because we are recording the audio and analyzing the audio, we have the ability immediately to determine whether or not that is an authentic um uh uh um uh uh piece of audio. So that's piece one, and that's hard to do in real time if you're a human talking to an expert. It's actually quite easy to do when an AI interviewer is talking to the expert because we have the benefit of being able to analyze that data in in real time. And of course, if they use it, the engagement shuts off. So that's kind of piece one. And then piece two is um we've built a lot of uh safeguards on the back end. So once we capture all that audio, right, we have lots of checks.
00:29:14 · Sagar Kadakia, Brett Caughran
We we we still have humans in the loop that that's still very important right now, right? Where the most important thing is the quality of the data that we're passing back to this person is authentic. It's high, it's high quality. And so we have humans and models that are reviewing all this uh uh for those um against those, you know, uh, against those checks. And so that's how we kind of do it today to prevent to bet to prevent fraud. [00:29:42] S1: What about um MMPI? I mean, this has been sort of a massive issue in the hedge fund world with expert networks over the years, uh sort of for a period of time, a flu sort of a fluid and shifting regulatory environment. Um and one of the one of the um one of the sort of serious footfalls doing these calls is if you sort of uncover MMPI and then that stock gets put on the restricted list across a firm, uh it's the right thing to do, but it's sort of people are annoyed with you. How do you keep these AI systems from accessing information that would put a stock on the restricted list? Or even worse, sort of not flag that as MMPI, and then that firm transacts transacts on that piece of information, um, which is a real problem.
00:30:27 · Sagar Kadakia
[00:30:27] S0: Yeah, that's a good question. Um, just from a compliance perspective, um when you build these businesses, that's like day one uh uh type of stuff. Um and I had been lucky to work in the space for quite quite some time, and so had been fortunate to you know work with a lot of compliance officers right before we kind of started this business. And so it was uh very important that not just from an MMPI and PII perspective, but every firm has their own compliance policies. Um and it was an attestations that they want to make, right? And and their own terms and conditions and what they like to uh uh ensure that their analysts are following and what they allow and don't allow, right? So no government employees, no medical doctors, no XYZ. And so these are all a part of this whole compliant kind of compliance posture, right? With MNPI, PII maybe being the most important thing. So um I want to start from that perspective, which is um everything that we have built here is very much through all that lens, right? From an MMPI and PII perspective, today we still have 100% human and loop review of every transcript that comes through for MMPI and PII, right? So we have a lot of operating procedures on hey, here's what we consider MMPI PII, and every transcript is being reviewed. If we identify it, most of the time we throw it out, sometimes we redact it. We're quite conservative here, and so often we will um toss the transcript because we don't want to change the meaning by by redacting anything, right? That that could have another impact in itself. So that's how we do uh uh the MMPI PII review. Obviously, we have a lot of models as well, right, that are assisting us um in reviewing. Um and so that's how we tackle that piece. And then just in general, the way that you have to build these firms is whatever your compliance policies and posture is, you have to be able to migrate all those policies into an application and then systematically enforce them, right? And so what does that mean? How are other ways that we prevent MNPI, PII? We work with a lot of funds where when one of their analysts um uh submits a question, our compliance team is gonna review those questions first, and then their compliance team is gonna review the questions.
00:32:36 · Sagar Kadakia
That's the first way to prevent MNPI and PII is to make sure that the questions themselves are not going to generate that level of information. And so we've kind of thought about it from uh every angle, right? From question review to response review to the overall compliance posture. What are the attestations, right, that and the T's and C's that someone has to sign off on, right? We're very clear, right, in all of the studies that we do. This is the information that we do not want. Do not give us this information, right? We don't want to know uh where you work, we don't want to know um any information that you feel is is private to your organization. And so those are all of the um uh steps that we have ultimately taken beyond just the review to ultimately ultimately prevent uh MMPI and PI. And we've done a really good job. Less than 1% of transcripts here contain that information. It's because of all the preventative measures that we have in place.