The Zero-Partners VC - AI Native VC running on AI Agents

Years ago, I opened with a partner that was also has a PhD in machine learning. We opened the VC, and we wanted to be a 2 people VC that works like 50. So we decided we're going to have AI engines instead of employees and let the AI engines do everything. And we coded everything ourselves from scratch. I will not touch everything. I will add that the most interesting part is actually how do you do the investments. As investors are 2 parts of the investment. The first 1 is I want to find the deals. I want to look at a lot of deals and decide whether I should even meet them. This is called the sourcing. So we built this system that basically, you can see here there are many, many. It source across as many, many, different, sources on the on the Internet. And then it looks it collects information and looks on a very shallow way how the where does it fit our thesis or team signal program. We have some criteria, and then we can actually have a pipeline of the things that came in. Once I I can look at, oh, this is interesting, I can click and say, okay. Let's do a quick screen. This looks interesting on the first look. The quick screen will run, and this is only based on stuff we collected. So we don't have a deck. You know, usually, it's a website. It's a name or whatnot. And this is an example that I run just before, so maybe I would get an email very, very soon in here as well. But I basically see, you know, this is potentially relevant. It says that the test is the mode. You know, the team gives me some information that I can look in. I can look at the LinkedIn, you know, investment scorecard. It already creates the scorecard, and it will highlight the gaps or the red flags, you know, something like that. Green flags, red flags. What should I add? You know? What should requires a verification? So this is the second, you know, super cheap 1. And spoiler, we broke everything into skills, and then we have a bunch of stuff. But because it is skills, I can also actually take some PDF like that that I found on the Internet that is, like, basically plug and play. It's a an accelerator with a long list, a table of long list of deals. I just send it to Claude, and I tell it, use your skills to extract the Martí up data and check relevancy check for ellipsis. It run the PDF content. You know, it's Kindle 63. Only 2 are relevant. These are 4 potentially relevant. I have here a a a whole analysis that explains. Okay. That's great. Once we have that, we actually built a system that takes a a a a now we we automatically, from that, reach out to the founders, tell them, hey. We really like you. Can you share a deck? We share a deck. Now we run deeper analysis based on a deck. So, basically, the system so so far, you don't need any human. You got the pitch deck. We already cleared this scorecard for the investment, and we basically start from that, and then we do a loop. We basically start to identify red flags, and, you know, issues, top stresses. You can actually see beneath it, you know, for if there is a full report on ethnicity, you know, the resources and totally so everything is, like, actually verified, not verified. Seems crazy. I already have a deal memo from the very beginning. And, again, I have a loop all the time. This is the deal in Richmond, you know, yeah. I have team assessment. I can see, you know, the rating of the team. All of these are right. And, you know, recommendation, you know, key risks and whatever, and I've a a a Martí analysis. We also have, like, you know, we created the thing that we can actually talk to the deal and ask questions and compel stuff if we had something similar. And like I said, it's all based on a a a bunch of skills. Right? So we have due diligence, scouting, CRM, schedule skills that you know that that work on on all of this. And the key idea, key insights that we had is that we basically every time we do that, we find gaps and whatnot. Do we have conviction to say no? We say no automatically with a nice explanation. If we say yes, okay. We move forward over there. If we don't know, we highlight the gaps. We we have another meeting, and every interaction, every meeting, every email, we find it. We push it in, and we run the entire thing altogether. So this is, like, you know, very, very quickly, what we have, some stuff I wanted to show. Another thing that actually we added recently, is the investment committee. So this is a simulated investment committee. 5 legendary seed investors debate the deal using all available. So, basically, we have a we said, why are we only in our own deals? Right? So we have a a a basically 5 of the best. We we look toward the best early stage investors. We created the profile. There's so much they explain how they do it, and then there is a guy that is deleted every time you choose someone else. And then they, each 1 says what they see, and then they debate, and then they they challenge for that. We add that as well. So Sorry. Can I just ask you to give them specific personalities or skills? Oh, no. We gave them. These are like. These are like real people that actually there is so much information that explain how they look at this. So we created Okay. Agents that actually represent them. Yeah. Yeah. I think some interesting stuff. Yeah. I wanted to say some insight. Right? So so the cool thing, I'd like to be like what capacity did. It's completely stateless. Right? So we have Müller with data and you run on it. So actually makes everything more simple, more robust if you can do that. We actually we are hacked because the problem we use skills, skills are skills that don't move. You need to install locally. So we actually have an MCP, and the MCP actually looks at the skills. So we update the skills in in GitHub. Then we SIM link into a a a place where we run cloud. We have a cloud instance in the cloud. The MCP knows all the disks, so anything that can connect to an MCP actually now connects and then we just like create a cloud code, sorry, instance for every skin run. So now we actually solve the problem for meter of probability. Anyhow, yeah, you know, you can you can, a byte code, everything, but, like, you know, kind of deleted things. Like, how do you make sure that there is no security? How do you run, you know, every week we run, hey. Analyze what is taking the most, coins or the biggest cost. We found out that actually generating doc x is more expensive than the deep Robert. Okay. Let's not do that. So I don't know. Yeah. I'm just close, the the yeah. Yeah. I'm timing. That's right. So so any questions? Yeah. Did you backtest this? Backtest? Like, did you write on, like, what information was available on Facebook or Uber or Oh, yeah. So so we we have multiple models. So we we we have 2 risks. 1 is the model hallucinating. The other 1 that the the the the the founders are bullshitting and not to be too naive. So, I I can actually share, you know, for the model, I'm just making, you know, we have a bunch of our first telly to to say I don't know. So we don't force it into an answer. Then you need to verify to say verified, plausible, or, you know, very suspicious. Provide sources for everything. We run different models from different providers on stuff We see that we get the same thing and we do it. So we have a bunch of the. All we run the same model multiple times on the same sheet to see that it comes from the same data. We use only JSON, so we don't use the pros. We use JSON to run, and we run, like, multiple different skills on the scenario. So it would be nice that you can write in the past with a previous deck of a winner. Oh, yeah. We did. Yeah. We did a lot of skills. Yeah. Yeah. Yeah. We learned that. Yeah. I think I think. Yeah. Alright. 1 last question. It's not open source, is it? Not yet. But, yeah, I think that we are all commoditized, so it doesn't matter. Yeah. Alright. That's another question. So human factor plays a big role in investing. So investing actually in the team, especially early stage, and all these tech and the hard coded factors don't really matter as much. So how do you so you only focus on super steady and upwards? We actually do super, super early stage. We do, first check. But there is a lot of there is a lot of actually collecting a lot of data, looking at the data, and being very, following how you should look at stuff. And the agents do that Müller, much better than any human does. We actually thought, what do you need humans for? And then we have answers for that. But, yeah, I actually think that there's a lot of, reading information and, none of them have any capital discussion. There are still 2 general partners in somewhere.