Founder OS #3: AI for Venture Funds: Playbooks That Ship

mindset I'm Alex project about connecting ideas, people together, build a couple of uh laboratory as we call it. One of our project is this space ecosystem. It's work society at least we will try to develop it. And we connecting people on creating some products and this sprint is dedicated to founder operation system is stack of tools that uh we can use as a founders as a entrepreneurs to build our ecosystems around us in business in real life and family and some productivity task. One more time my name is Ray. I'm N Madrid today's session mindset and trying to put a lot of different folks really create interesting folks together and uh this initiative is basically one of the sprint of founder operation system that is why we got so many founders and this is the first session around this and I believe it should be quite good because today we have three speakers and basically they will showcase how they use AI in their day-to-day life in venture funds uh And it will be three of us. The first one will be Euro. Uh he's uh head of scouting in uh Ultra VC who is and Stan Gaggay he will speak about how he use AI to reach more investors and not only but uh he basically good at it. Stean who is also with us today uh he will talk through his lens and it's very interesting because Yuri and Stan the lens more like web two focus uh and uh stipance is more like web three so it's different words but at the same time it's the same deals and uh everybody is uh interesting in AI right now so we will have three speeches it will be like 15 minutes uh from 15 to 20 minutes if you will have time in this 20 minutes you can ask some questions otherwise it will be more on the end of the 20 minutes we will gather more like round tables with all the guys and uh you will have opportunity to ask more questions if you have so yeah we will have uh recordings but uh one more time it will be not uh straightforward so it will be better to just sit here and listen and

ask questions uh online so I don't want to catch so much time uh from the guys actually because uh I'm very interested in this topic as well and I want to ask Euro maybe to unmute him. — So you are passing the mic to me right? — Definitely you're the studio. — Okay Fman. Hey everyone, happy to see you all. Actually, we will stand our bit like in the shock like since yesterday cuz uh like it was planned to be a like cozy event for like 10 like 12 folks like discussing some stuff and now we're seeing like 65 like it's crazy like uh there's lots of interest in the topic. Uh please like quick disclaimer was planned to be like a cozy where we sh like uh all of our like back end sites maybe it's even better I don't know but uh me personally I do not have like any like prepared like well prepared speech like for 60 something people so uh actually alced and seed VC we invest uh in AI native companies with potential to become full stock a companies across Europe have some uh deals in the US as well. We've started actually like a few years ago like two something years ago with a thesis like initial thesis was to back uh impact driven startup companies uh and not only to back them with pre tickets but also to uh support them with uh some kind of like help some kind of mentorship because we are not big fans although we invest pretty early on we are not big fans of spray and pray approach when you invest lots of tickets and keep your fingers crossed that something will like kind of play out. It means that we do I think a bit fewer deals comparing us to typical like active US-based preede VCs still believe that we are like in the top quart in terms like comparing us to like European colleagues. So it's kind of 20 something deals per year and it allows us to be way more helpful to the founders because uh of course you can't find too many VCs who will tell you that they are not smart money but we are very strict with ourselves what we can bring to the table except the checks. Quite often we even skip deals when it's sent to it we cannot be anything that just asset figures on the cap table. So we operate used to operate like on a cohort basis. So comparable to like accelerate accelerator model uh now we're figuring out like uh so maybe we'll do some kind of a mixture approach but uh uh like yeah as I've said doing like making like 20 something deals per year and I'm responsible for like scouting I'm responsible for analytics stuff at Ultra. So basically I'm the person like my main responsibility is to always have great founders around to invest in and support the current status of Altra is uh we I believe we know like by now like we know how to generate enough like pipeline in terms of quantity we screen around 2. 5,000 companies per year so you can like calculate like success rate the success rate or like the investment rate so it's like lower than 1%. So uh and do like 15 to like 20 something deals per year and uh when you see uh look through so many leads so many companies it's just like a matter of probability theory to like find some great anomalies. So basically I believe we are uh doing pretty good job in terms of generating like enough pipeline in terms of coin as of now and when we've started thinking about like uh okay so we invest like in fullstack AI companies but where should we apply for like problems uh where we can that we can solve with AI power tools found out that the like real room for growth is to for us is to like and it's an like never- ending game actually as of the same for the all of this is to learn how to spot like the best deals and maybe the deals that are hidden from like the public eye and learn uh so um maybe uh it like case of utilizing some kind of AI tools uh and automations across like our VC colleagues to just like widen the top of the funnel like the number of companies that are screened. Yeah. But uh for us uh the we believe that mostly the major focus should lie here. So not on the whitening like this, not only on whitening like this top of the funnel, but also in learning how to spot like this hidden deals uh that are closed like without like even going public. And uh there are like a few ways to approach this. There is of course like this traditional way and uh it's like a VC business. So develop relationships with like top funds with funds that are uh often do common deals with tier one funds and uh figure out the way uh like figure out those we have like most like intersections in terms of the CS Swiss and develop like relationships and uh like great deals and hope that they will invite us. So it's like a typical like VC business. But uh getting to like the AI stuff, there are also like two ways to approach like the problem of finding these companies that are hidden from the public eye. And I call it like innovative intensive and innovative extensive. So let's start with the first one. So basically innovative intensive is that learn how to spot like most promising founders early on. So you can think of hypothesis like uh x top geni companies employees that have recently left the companies and uh launched their uh own startups or uh sales founders or like spotting the signals from top VCs. Innovative extensive is like as I mentioned just widen the top of the funnel. So learn how to screen not like 2. 5k uh startups per year but 10k 25k 100k companies a year and uh you can think of hypothesis like AI prescreening assistance or AI intra call makers uh today I'll focus more on the first one on the innovative intensive and briefly tell you what we are like building in extensive as well uh and I believe uh like I hope that this uh like meetings, these gatherings will become like regular sometime in a few weeks we'll be ready to like showcase like with a proper like uh front end uh how does it work? So but uh let's start with innovative intensive and to remind you it's like uh learning how to spot like the most promising founders. Well, actually where we've uh came up to this uh problem uh the most like obvious thing to do is to go on market and try to find any existing solutions that are on there and uh actually we've done that

uh we've demoed like uh a few of them and maybe I can share my screen and share you unfortunately like uh I'm not allowed like to uh share publicly like user interfaces etc. But at least I share with you like the websites and uh we'll tell you a bit more about like the insights we get from these platforms but I believe like for other it maybe uh may make sense to at least like request like a demo and to try this platform. So basically we found like three of them like two of them are more like maybe about pretty much the same thing and this thing is that they uh parse lots of data they parse like um crunch base like LinkedIn lots of media stuff etc try to have like as like relevant I try to update like the database like the data state very regularly like harmonic state they're updating their database like uh every two weeks Spectre says every month. So uh basically they try to parse every like parsible uh startup on the market and uh apply as many filters to them as possible help like this is to understand to like find most perfect matches u and they are pretty much comparable each other in casual in a way. So it's can be compared like a spectre is a legot that harmonic is more of a like mechanic like electronic constructor. So basically takes more time to board here but maybe there like the outcomes will be a bit better but uh actually they're pretty much comparable. Uh to be honest like after like demoing uh the platforms they are great in all of them. Uh but for our specific uh case and our case like our like bottleneck uh in scouting is not like not having enough leads uh enough companies to look for as yeah but like is to learn how to spot like the hidden one. I'd say that uh most of the functionality in both of uh extensive like is does not bring too much value to our specific case. I know the cases of the funds that are uh really benefiting from either harmonic or spectrum and usually it's the funds that have some kind of um strict restrictions in terms of their investment thesis. So they are either can invest they either can invest only in a specific geography for example or in a very like tight and narrow market and this funds usually have like less like companies on top of the funnel so less companies to screen. So that's why uh they have some so usually the bottleneck is like on the top of the funnel. So these platforms are like fully like solve their problem. But still uh I found two interesting features that are solving our specific uh problems and uh this uh most inspector by the way and these features are called like talent signals and people signals. So basically what talent signals is a feature where you can pre-build like a list of uh companies you want you want to monitor like in our case for example it's like top uh companies that are building foundational like models and uh this uh and the tool proactively monitors like the social media profile and not only the social media profiles and the media as well uh of their employees and uh spot those who have like just recently left the company and

uh stated that they are starting to build their own stuff. The hypothesis here is that people that are working like in the frontier have like experience of working frontier uh technology they have uh deeper insights uh or maybe not but like the uh the probability is like higher. Also the hypothesis is that people with such kind of experience they uh it will be easier for them to raise uh at least the first maybe the second round cuz uh like for example if you want to build a company like building some kind of foundational AI stuff right now the competition is pretty fierce and uh the ex top like employees of gener companies have some kind uh unfair advantage is uh called this way. So there is a very interesting feature like with this talent signal. So where you can just like put in a list of or like companies you want to track. Uh then it's proactively uh sends you some calls when uh this or that person have stated that they left the company and uh launched their own startup. And the second thing uh is also pretty interesting. It's called people signals. Uh more or even like investor signals uh and it's more about um yeah it's called investor signals uh spectra. So it's more about so you can like set up a list of uh investors that like may uh have or are having like a you have like pretty good overlap in terms of like investment and then this uh tool proactively uh monitors uh and uh sends signals if like a few of them are interacting somehow. how with employees or founders of this or that startup. So think of like adding on LinkedIn like uh or like X Twitter X uh liking their posts maybe commenting the post etc etc. Usually it can be a pretty strong signal that they are in the middle of uh like discussing like the deal with star and sometimes like the hidden sometimes you don't even know that the folks are like building a startup or the startup may be in a st mode but actually it say the like signal that like some of the top tier receives that are focusing on preede are like interacting with uh this uh founders of this startup and maybe make sense to like write and if you see that they're building something that maybe of fit with the strategy maybe it makes sense to reach out to them and uh ask whether like they're raising or not and maybe try to get some allocation this route as well. Also there is a feature like with like and there is a current hypothesis we are testing with founders that is building the styles mode. Uh so of now founders that are stating that they are building installs is we are considering it is a meat like uh mid power like signal that the company may be building something interesting because uh at least it signals that the founders are a bit aware of the like VC landscape and trends uh and they are aware that this to this approach may be used to spark some interest from VCs that they kind uh we are pretty seat we are so secretish so we're won't tell anyone and unless they ask us that we are building the startup so it's like the value and the strength of this signal that the founder like go tells will uh inflate uh but for now it's still like a me power signal that the company may be building something interesting so uh figuring out uh like finding these founders and figuring out the way how to learn what they building is a pretty interesting hypothesis uh to be honest still much to be done here not only by us but on market as well. I was approached by a few companies that and a few startups that are claiming that they can and they are already like able to predict what kind of thing this or that styles founder is building. uh but I've tested a few products and uh at best they are telling uh something kind of this fog is building in the education space or so. So not so much value and as of now the problem is not

solved but there is a potential to I believe that there is some kind of way to build this prediction model as well. So that's uh regarding our like innovative like intens intensive hypothesis. Uh so as you see we are not uh like super actively building our own proprietary tech here as of now because we've like just research mode but we've tested like a few things and I believe like these talent signals like investor signals and sales founders hypothes are pretty strong. I'm not sure uh whether we'll buy one of these products or not like uh in the near future, but at least we'll consider the ways how to um uh maybe implement some kind of uh our own tech to solve this. So we'll see getting to this like innovative like extensive approach like the approach like how to learn how to screen like not like two and a half thousand companies per year but like six times more. So uh here we have uh a bit more uh progress like in terms of our internal tech almost already built like uh like at the model level it works but uh we do still do not have like the front end. So unfortunately I go like walk you through today like but we've built like AI pre-screening assistant our own propritor one the technical name of uh the assistant is Midas uh like we believe it will bring us uh like fortune making some fortunes basically Midas is able to do is to we feed it with a like a pitch deck and ideally the uh application I forgot to mention And uh we have pretty so if you go to our website you see that we already have like this applied now thing and where we have like a pretty simple type form uh so we gather like some kind of inbound applications pretty like tangible amount of them. So especially uh in the like active fundraising uh season. So basically u uh in the best case scenario we can feed like midas with like the pitch deck and this like website application and the outcome uh that we are were targeting is uh to like rank all these applications uh how likely we'll be interested to close the deal with this or that company. And uh of course there is like some basic criter criteria like domain, geography like round size uh traction numbers or whatever. And to be fully honest this level of uh tech I believe that uh it's like the maximum capability uh of like such kind of assistance. So they still fail miserably when you ask them to perform some kind of deeper analysis. So applications I've like depersonalized this you can't like see the names of the companies but actually what we did we like uh ranked them manually and also we ranked them through medas. We have do not have uh so much time to uh walk through all of these like criteria but you can see that the uh many of them. Basically our major goal was to um have as less false negative outcomes as possible prefers to skip like it has some kind of potential. Uh we believe this like false negative outcomes will be less than 3% it will be a success. current state it's around like 6 7% so we are on the way but still like it's already like providing some kind of uh value so it's closer like to being like a junior ad so yeah in a nutshell uh

we so in terms of like training of the which we already have we like we have developed like this assistant and uh we'll be happy to walk you through like the whole front end thing like maybe at the next time. And in terms of like some hidden deals like this hypothesis with talent signals and investor signals in my opinion are pretty promising. — VC funds are making the same like founders still like founder. So you're just looking what's going on around on this uh B2 VC tech and stealing the ideas. Uh I'm not sure it's about stealing the ideas actually because um like for the platforms actually it's not uh what they selling their ICP is not like VC funds like us. Their ICP is VC funds of a more of a narrow like uh focus and uh they are selling to the funds that have like difficulties with kind of the top of the funnel. Uh I'm not sure it's about like stealing the ideas cuz maybe we'll end up like buying from them. But uh it's how like economy works. So if you see that like it's cheaper for you or economic more economically efficient for you to build it yourself uh I'll be the first one who will share like this feedback with the folks because I believe it will allow them to like like adjust the strategies because uh because it makes no sense to pretend that you know uh anyone is so kind that will buy just because of a demo. uh but uh it really makes sense to give like honest feedback. Uh this is the only way how these folks include included will be able to build some kind of huge business and they're aiming to build huge business actually. Um I believe we have a pretty big room for for the next event as well and uh for the round table in the end of this meeting uh as but uh let's maybe go to the uh second speaker because uh Stan promised me that he provide some demo on this event. Uh crossing my fingers for this please I want to see something. — Hi everyone. Uh yep I've uh put something together for you. Uh before the meeting I had like couple of hours of time to build something. Uh I will share my screen and show it to you. So just couple of sentences about myself. I'm also a colleague of Yuri. I work in ultravc for a little bit more than a year and I'm responsible mostly for uh VC engagement right now also for some LP communications and very recently also for doing some outreach to founders. I was really excited about the genai space for some quite some time. Uh so my skills were useful in some other areas apart from uh finding investors. Uh I

had a chance to put together a very simple NA10 workflow because NA10 is my favorite tool. I think NA10 is the best. Um apart from that we will use some help of Linked helper. It's a tool which allows you to scrape LinkedIn also to put a list of LinkedIn uh leads and then it can outreach uh for you. For uh for the reason of saving time, I have prepared already a list which I downloaded from linked helper. To be honest, you can find uh this kind of list from uh some other tools as well. And one of the tool is API is it is another great tool which I will be talking to today about. They have a store of uh APIs. For some reason, it's not loading. Let's see. And you can find different crawlers, different scrapers here. You can find of course the one for LinkedIn. So here's some example of how you can do it for uh let's find something what can scrape LinkedIn search. Uh we need profile scraper. Yeah, something like that. Okay. Actually, I already used it today. Uh and I was looking for exited founders. So um you could look for exited founders for different reasons. For example, you want to uh sell them your startup because you think they might uh like maybe could be an angel investor in your startup or they can invest in your firm because they could be potential LP or you can find them because you believe that you want to keep them close to you because they will start the next company and you will invest in that company. So for us as a VC exited founders could be a three-fold useful contact. So uh for example you can scrape United Kingdom here for uh exited founders. I'm putting it in quotation because if you put it in quotation it will be looking for the exact uh search there is build there could you can export uh you can export different uh CSVs and it will be provide you very big list of uh information about that person because this scrapes LinkedIn of course you'll get information about this person how many connections they have where they worked and all of that good stuff. So going back to our first step, you need to find some list. So I already told you could use link helper or you can use API to find your first list of uh LinkedIn. So with that you can use something like API enrichment that uh also you can find it uh on API store, LinkedIn profile search scraper. to be honest like if you put some keywords into API you will find maybe 10 of them which do the same is profile search let me find the right one so right now we need it to enrich our profiles so something like this LinkedIn profile details scraper yeah uh here you can also put uh put them manually for example bulk edited and just paste from your CSV a bunch of LinkedIn in the and you will have enriched profiles. So here's the result. I got the list of enriched profiles. Um and then you can start with an intent. So the reason I'm doing some of the steps manually because it's just faster. you kind of can um navigate it easier if you're doing some steps manually uh because API is uh doing something in the shadow and uh you cannot control the the list as and also for the reasons of MVP it's better to start on the small list uh of like 100 items then look how it works and then scale it with APIs uh much more so like 2,000 uh LinkedIn profiles. So here is uh just an example of how APIs work in NA10. You put your URL, you put method post, and you put JSON who you want to find. So here is uh the example of who I'm looking for. I'm looking for exited founders uh from Germany. This is just an example of a note which can find uh people. And if I wanted to do everything uh automatically, for example, every week I want to scrape uh exited founders from Germany, I could just put it somewhere here and ask my NAT workflow to run every week and every week I will get a fresh list and then I will put them through my uh workflow. So you can see here on the right the list of items uh and the list of profiles which we received. So there is uh hund of them hund of people and you see uh how many interesting information we have about them. So everything from our DSV. So here you have everything. After that after you scraped uh and put all your data into uh into an it loop and what the loop does it uh um edits field it creates a profile AI. It creates a nice uh way for us to create. So here what I've done I created a small description of a person which we are trying to analyze. So here uh is a short description of a founder which I think has exited their previous startup and

there I take some variables from our CSV and I'm putting them here and you can see there is a lot of variables. There is information about uh their work experience also what they write about themselves on LinkedIn their headline their first name all of that good stuff. So now when we created a profile which we can put into a prompt I'm using u open AI API to send it and to ask um open AI to kind of rate this person to say hey is this LinkedIn really fits our criteria or not do we really want to talk to this person or not if he's actually an exited founder for example so this is called structured outputs this is a very powerful ful thing. It allows you to create um prompts which will give you in the end a JSON which then you can parse and you will always get the same structure. This is very useful. Uh so you can read through the prompt but essentially what it says hey here is u uh you are a structured data extraction expert you work for Ultra VC we are invest in early stage AI native companies um and uh you are provided with a supposedly uh exited found you need to help to find if this person is actually exited and create a personalized message for the further outreach. So this is in the structure of JSON. So structured output there are different uh questions I'm asking separately the uh llm I'm asking for example exited founder uh return me yes or something else and the description is uh so I want you to indicate whether this person has previously sold a company. So uh the reason for that is because in the LinkedIn you could have people who are just writing about their experiences. They might not be exited founder but they have worked with exited founder as a personal assistant or something like that. But this uh this tool will actually read the context and it will understand the context and say yes this person actually exited the company he talks about it in their LinkedIn profile. The same goes for other things. So is this person worked with AI for example? Is this person sold a startup or is this sold an agency or this person even like interested in startups? So a lots of different flags you can use for your purposes. If you're for example fundraising as a startup, you can put is this an angel investor or is this angel investor invested in B2C. Uh a lot of things you can do with this and create as many flags as you want. Uh another thing you can create fields is output as a string. So you can ask some things like hey create a message uh based on this profile. So here I'm asking the LLM to create a personalized 300 symbol messageing his experience and I want you to ask him to collab with Ultra Vistle. So I don't have to write those messages uh personalized messages but I believe you understand uh how it works. Now we parse the replies because the answer from uh open AI comes in this format. It is also JSON but it's kind of a content blob of text and you need to extract uh the right fields in the right uh format. So in the end you get this exited founder yes AI native something else uh startup enthusiast yes and you get the message and the description of a person and then it all goes and get stored in a in your Google sheets. So I can right now show you how it's being stored in the output section. So uh as you see like the loop is going and it adds more and more uh profiles to the uh to the Google sheet here as you can see the description they say hey this person is Paul an exited founder who successfully launched and sold business in the property financial services and pet product sector. You can also run another uh similar workflow and for example say I want only people who have sold companies in this in the B2B B2C space for example in fintex anything. So you could add even more flags but it's just example. So AI native something else. Yeah probably not because uh yeah I don't know maybe exit founder yes because it clearly states that he sold something. Startup enthusiast yes message. Hi Paul I'm impressed. uh by your journey from uh military leadership to successful entrepreneur especially with Arc X and Safeex. Your ability to invent consumer tech and pet products is inspiring. Yeah, let's connect. And so it's pretty um pretty generic but it's it mentions uh the person, it mentions his companies and uh there is clear call to action. So of course you can improve it uh many folds but uh this is a good start. It's better than generic. Uh okay. So yeah uh this is the end but uh if you want to improve it and how uh we would improve it if something like this would be uh implemented in our uh like hypothesis testing for founders for LPS or for DCS. The first step would

be change the trigger. It would be a chrome job which will run maybe every week or every month. Then I would add uh APIs which will do all automatically scrape for the uh requirements of our hypothesis and in the end very likely we would uh assign a owner of the person who needs to outreach to that person and uh connect CRM. in CRM it will put hey this person is responsible he should uh send a message and also would be nice is to create a CSV from uh all the list for example we assign uh 50 people to uh a certain person and he can put that CSC and put it in something like LinkedIn helper or phantom buster and then outreach them all automatically so to save time. Yep. I think that's uh that's it. — S or Sam. — Yeah. — Stan is what Stan is doing is helping the founders in their batch and the VC fund to find co-investors or somebody who can help them. So he uses it actually like every day uh in his work and so if you're a founder just mention that to quite good at warm helping for every founder uh but uh maybe we can uh find out what's going on in other side of VC and with Stean Guni who is actually making kind of the same but in another Maybe it's less linkad in it. Maybe it's more like kicks uh far caster or any other tools. It's interesting how he uses one. — Cool. Yeah. So, it's a bit interesting because we had a system uh I mean we still have the system uh that uh runs kind of AI automation for the fund and we had a bit of a disaster this morning. um we had to like restart the parsing. So I won't be able to show the like it's been working like I showed it yesterday like there's a few people on with whom I had like uh one-on-one conversations and show before but unfortunately I cannot show you the live demo right now because we are doing reparssing and we have like 11 million uh accounts uh to to run again uh because of technical mistake but I can just give like a high level overview. So let me talk about like what we do, how we do it and what is the kind of longer term plan and where I think uh um what is the AI native uh investment uh will look like in a couple of years. So I'm with Cyber Fund. Cyber Fund is a proprietary capital investment firm. Um so we invest um uh about um the fund size is uh around $200 million. We invest in early stage companies uh primarily in AI and also in blockchain space. Um we invest preed seed and series A check sizes between half a million and $15 million per deal. Uh we try to be kind of researchdriven and invest in um like identify thesis, do deep research, identify the best founders and the best teams in that space. uh sometimes lead around sometimes follow on uh we also run accelerator we run a bunch of conferences there's an open source AI conference in two weeks in San Francisco that I'm running if anyone is interested DM me I'll send you promo code uh would be cool to me and um yeah so I think first of all super interesting uh demos and uh super interesting uh conversation so far this is kind of zeitgeist uh in a sense that most uh venture funds, most hedge funds, most investment or that are you know trying to stay competitive and win are experimenting and building different products or using different products to augment their work with AI and I think it's just starting. I think the capabilities will become much uh better and it will be much more efficient. For us personally, I've been looking at uh different tools like harmonic and and spectrum and others. Um what we decided to do is actually to build everything internally. Uh why? Because I think this is um one of the core capabilities like I think in five years I want us an AI native fund. So in the sense that first of all we use PI to

automate internal processes like finance and generating memos and generating communications like recording every call uh writing all the operations running all the uh portfolio support and and generally sharing data and we're actually incubating one startup right now that is kind of like an operating system for companies uh specifically starting with investment organizations that allow you to automate internal processes not as much discuss discovery and lead generation more like operations. So instead of having operations team, you just have a shared database where you have all the data, all the calls, all the to-do lists, all your financial information, all the memos um all the information and then you can ask uh you can build workflows for finance, for marketing, for accounting, for legal, for whatever. Um so in terms of how it works uh for us for like uh we started Bill we started with uh focusing on deal flow generation and I think the goal for this system is even not just like we still have majority of the deal flow coming from the real physical network but uh what it helps with is understanding the market dynamics and market environment. And so far, just to give you some numbers, uh we have uh about 12 million different uh accounts and data sources track that we parse every day. Um and we use that data to um uh basically through the list of filters to identify and come up with interesting investment opportunities or interesting projects. if those are open source pro. So far we work with a few data sources and for each data source there's a different algorithm how exactly this parsed for example one of them is GitHub. So GitHub is a uh kind of uh open source repository platform platform for open source projects. What we do we have a list of interesting relevant projects. the best open-source AI projects you can use if you go to GitHub API there is something called stargazer you can find the people who were uh putting star on the specific repository then you can find and filter for the people who across these uh top projects and also were among the first to put the star and this is our seed that we then use to enhance and basically track those people where do they put star so these people are early discoverers of the interesting projects And then for each project uh we kind of increase the the the database of seeds. So what it gives at the end it gives you early signals on the growing opensource projects. Those are typically not companies you cannot invest in them but first it helps you understand like what where is actually the attention like right now everybody is talking about reinforcement learning environments. uh maybe interesting evaluation uh evolves for AI or maybe interesting agentic applications or frameworks. Um and then we track each one of those uh based on the filters. Is it a company? Is there a business model? Is it investable? Like um uh is it a research project? And it gives me a weekly uh kind of overview or weekly uh digest of the most interesting project from GitHub. Similar thing we have for discord servers. Um so there's a bunch there's like a few hundred discord servers uh from different AI communities. We just manually set up some channels to follow. Uh same can be done with telegram. I just don't find in telegram. It's typically just reposts from other networks. Um and then um every every message uh it is tracked by the number of uh replies or

emojis. basically how much signal, how much traction does that get and then it's filter if this is a link to the company or link to the project that somebody is building or if this is a uh some useful information about the project that we already have in the data. The most useful the most interesting data source I would say Twitter on Twitter we have like most of it is from like we have I think 5 million accounts at this point. these accounts it started like how we start we started we we just manually selected 1,000 uh top investors angel investors and venture capital kind of firm uh employees or partners um who invest in AI and then uh we started looking at um who are they following so it's the same idea I think it was already discussed today so if you track people on who they follow recently you can increase this database and there's also a notion of score. So for example, there's like a maximum score that Mark Andre or or Andre Karpati or uh Clam face clam has and then um depending on how strong the connection is, how many degrees of separation uh we can assign different score level to the people that they follow and then this is updated based on uh how like if that leads to any interesting uh leads. So we have like basically two kind of concepts within we have seeds and that is like right now uh four like almost five million accounts. So those are people that we track every day what they do and then there is um uh leads. Leads are other accounts but those leads are going through the bunch of filters. So far there's three-step filtering pro. Step number one it is uh we look at each account and we try to say is it a company? Is it a startup? Are they early stage? Is this account been created in the past one or two years? And if this account is um active like is it is it an AI related early stage company or project then there is and this is done like there's no LLM there's no like uh it's purely keywords and and looking at metrics and data that you get from API. Then the second step is we use like a cheaper model to try to understand if this is a product is it actually starts up is it um uh is it doing something um so it's not just keyword but actually running a smaller model llama 4 to understand if it is uh um an AI project if it's uh satisfies the investment criteria the investment piece that's just a pro like similar to the uh to the demo that uh that we just seen and then it starts writing the memo. So for every company that goes to step two, there's already a memo. So it goes it treats everything on the website. So there's a playright kind of uh headless browser that checks everything on their Twitter account, all the posts. It checks everything from their website. It tries to find the names of the founders. If it can identify it, goes to LinkedIn, scrapes the data from from LinkedIn page and that information is added. And then there's this third tier, the final uh filtering step where it is looking for um basically perfect match. And to give you an example, like last run was around 900,000 new accounts that were found and there's only 300 candidates that was uh that went to the last step. So it's like it has to be super strict because you cannot realistically manually check like you know thousands of of potential candidates every week. You need to be pretty strict and comprehensive. So for each one of them we have a founding team their experience how much this company raised if there's any information including crunchbased uh data from API or or just open AI. I think it's uh GPT4 or five uh mini with web search that just does it like a quick search. it writes a memo and um I can go in manually in the readal and then and probably the most important thing that we have is once that happens there is an ability for every team member and our team is just three people so it's not a big uh big one but uh every team member can go and independently read all those memos and put a like on the start that they feel like it's can be a potentially interesting candidate and that is the beginning of the reinforcement learning loop because at the end of the day it is not just setting like setting up scrapers is easy. You can use hope AP5 perfect service super good for us we because of scale because like if you have like 10 million it would be just too expensive so you need to have like your own infrastructure is 10 times cheaper um maybe more um but that is kind of that's that information is available to everyone I think the unique advantage that you can build on top of that is to build this reinforcement learning loop where you actually try to like even uh like right now just Okay, is it basically a binary signal? Yes or no divided multiplied by three team members. But over time it can be some kind of like a um I mean right now we have granola. So all calls are recorded in granola. So we have a discussion just a regular call discussing a startup and somebody says okay like this is not good because like uh I don't know uh uh we don't invest in companies that uh in like in certain juris we don't invest in Asian companies because we don't understand that market. Okay. So that is recorded and that is another kind of information for the future. So, so basically after that you can use this py or something like this to optimize the prompt and to um improve the quality for for the next batches for indexing and then if there is a company that get three likes from every team member then we can backtrack basically like back propagation uh to see okay which of the leads which of the accounts actually influenced or were following this project how did it get into the database is and then we increase score for those accounts because those accounts are great at discovery and it's random like we don't know like who those people I have no idea like sometimes it's just like random unknown person with anime avatar and in most cases it's a known person with anime avatar because it's AI and the AI industry is apparently people watch a lot of anime so because it's preed a lot of information is not available what you can get you can read Twitter one thing that I'm kind of interested in one of the next apps is to do I guess uh I can talk a little bit about road map. So

what what are we planning to do next? First of all is multi-tendency. We want to do it not just for AI but for other verticals specifically web 3 and robotics which is different criteria different heristics different kind of filtering. Um I'm like I have this hypothesis that you can use uh scientific data like um archive or or or or uh basically a king movement uh of people in academia somebody who's leaving uh who has like a good citation rates or published uh prominent papers and leaving academia and starting a new startup kind of like World Labs that would be a good signal before they actually raise money. Uh obviously following relationships um and then uh intros because for every startup that you identify once you have a lead let's say so 300 that was like an outlier because it was just a huge batch last week on average it would be for us uh between 10 and 20 new leads every week. So for these uh let's say approximately 15 companies every week I want to know how am I connected because cold outreach it does work but it doesn't work as good uh in most cases it's just uh you have like there's some way to have a warming reduction and warming reactions work better if you can get it. So, uh, one like I don't know how exactly to do it, but one way to do it would be to have like a sandbox with the browser where I'm signed in my LinkedIn profile on behalf of my authenticated LinkedIn profile. I can go to the profile of the founder and see common connections. I don't use LinkedIn really almost ever. So, I don't like believe in LinkedIn. Yeah, it's there's a lot of noise on LinkedIn, but let's put it this way. Um, another way could be somehow highlighting the Twitter accounts of the funds that we partnered with like for example we invested in 20 different funds as LPS like fund of fund structure. So, so like through those we can get like maybe warmer connections. Yeah. So, so this is an open question and that's been kind of experiment that we've been running for the past two months. So it's like actively in the development and as you can see it doesn't work perfectly because yeah we had this uh this emergency I am actually happy like for those who are interested uh like it will probably take so the full parsing it would take like four for four days because again there's rate limits also we cannot use official Twitter API because it will cost like million dollars per months uh there are ways to like actually infrastructure is not expensive so for this scale Um, our infrastructure cost is about $500 a month. So, it's super cheap, including LLM token. Actually, we added GPT5 recent, so maybe it would be like I don't know, but it's within $1,000. Definitely like cheaper than, you know, a typical analyst for the venture fund. And yeah, and it's been like a new experiment running for the past two months. I think this will evolve super quickly. I think there's a lot of First of all, like where where's the alpha? like what you can do, how can you make a lot of money or uh um make something super interesting with this approach? Uh one is hedge fund strategies. So I've been talking to a couple of discretionary hedge funds that been basically building inhouse corser for trading. It's a pretty sophisticated tools um much more complex than what I'm describing. uh those tools are like tracking if you're investing like in public markets or if you're investing in crypto you can track like people who own significant percentage what do they do you can track other funds because their wallets or their accounts are public and or you can track like SACE information you can track uh news behind the news basically and that's been but hedge fund been doing it for like LLM not LLMs but NLP being a tool for hedge funds forever. So, it's not new, but LLMs were never as powerful as they are today. Let me maybe show just a quickly like how it looks because I think I I've been talking like I don't like talking. I like uh I like demo. Um so, this is the graph database that we have. Uh like this is an example. These are all the accounts like you can see the mill600 something uh different uh data points. Each data point is a an account. like this guy his co-founder somewhere I have no idea but we know that he's a person and this is also a person but some of these would be startups and then you can kind of zoom out like this is only 5,000 because not load 11 million my computer blow up uh but you can see there's like some clustering so we can uh use some interesting math so like this in new forj there's this concept of communities so it's basically like subgraphs that you can create and graph traversal algorithms so there's like a lot of complex math that might work might not work but it's interesting. Then uh this is the front end. It's extremely ugly. Uh why? Because nobody wants to do no nobody likes building front ends. Uh this is in process. So this is like the actual scanning in process right now. There's like h half a million of of of there none. Yeah, unfortunately there's none that that's already fully scanned because we launched a couple of hours ago. But you can see that example there is a uh I don't know like this is a fund for example for each of the company there's a bunch of information how it was discovered all the data from Twitter and then we have like this tier one tier two tier three filters that I was describing and at the end for those that will be the finalists there will be a report so there will be like a two pages memo uh generated based on all the information from Google from LinkedIn from their website from Twitter on founders and everything. Um and then there's like different filters like is it high signal firm? Is it a candidate for seed? Can we use this to discover new accounts? Is it a like all the information that we scraped from the website will go there and then it outputs the JSON and we also have an API internally. So I can actually qu like I use corsor I don't know if you can see it. uh maybe not but basically like I can use API to query and use it in my cursor to automate some of my um some of my work. Uh yeah, this is like great. Uh — yeah, maybe I can answer questions. So sorry for [ __ ] up with the demo like not intentional but uh super uh unfortunate. We hear not about TED talks. we hear more about the roity because nobody said it's easy. It's always like this and this is why we're here to share this experience together. Maybe we can just make short kind of like round table around this topic with Stan and Yuri and maybe we have a couple of questions so you can uh answer them uh ask them in chat uh it will be much more easier. So if you have them just uh ask. But the first question from my side will be uh for everybody here because everybody talked about this filtration of millions or hundreds or thousands different accounts on different platforms. What is the craziest uh interesting or just field in your database that signal you that this company is crazy good and you should invest in it but you didn't believe in it previously. — Yeah, good question. I can say that a lot of that comes from like for preede stage that comes from founder energy from a lot of interpersonal. I don't

believe there can be one just one characteristic that will just kind of uh turn from the upside down like the whole like picture on the company but uh talking about like a native and like fullstack AI companies. So what we found like what is really important like for us and what is a strong signal that their tech folks it can be either like a CTO with like the devs or maybe co they are genuinely like uh excited about like playing around with like LM stuff. Uh so they are doing it just for fun you know it's still a part of it is still like a kind of black box still have like to like iterate for some time with uh the tools with the fundamental models etc in order until you like find some kind of okayish like so and the chances you'll find like actually like are much higher when like the folks are just like natively enjoying uh like playing around with these tools. So uh it's pretty easy to spot like because like they are always like very excited also to tell that what they found out. They all always come to the calls with some kind of insights. This like framework or this like uh set of prompts and it worked you know or they always have some kind of internal jokes about some unexpected results. So uh but it can be a pretty strong signal that they are more likely to find something that works. Actually — I like the idea of Yuri. I think I could create some filter which would try to understand if there's genuine interest. It would be like interesting site uh but to be honest like it's hard to find something creative. A lot of VCs just check uh check the employment history. If you worked in open AI as a researcher right now, most likely a lot of VCs will be interested. So there are a lot of uh a lot of weight on employment history I believe because uh those build big companies they are also a filters for greatness and uh definitely we need to look for them. So I have also created different tools which scrape crunch base for the list of companies which look for excellence and then we find companies find founders which are have been there and now they are creating startups. — Okay. So we have a comment here as well that maybe students or alumni from top five tech school great signal — can be but uh based on our like experience on our track record like investing in students uh like we still do not have like we had a few attempts uh but uh no successful one but we'll see. uh but uh let's speak more about founders maybe and when I spoke with founders uh any like advice because it looks like founder right now should be a full-time influencer to catch your AI tools just to uh make a lot of connection uh telling everybody about his company if you uh scraping his LinkedIn or talking in Twitter so it's kind of like part-time job but uh to catch Not always. I'd say like uh one of the best paths to like get aligned with VCs is to come through the like warm intro. Like actually it does not take you to be an influencer, a public person to just like meet some people and build relationships with them and maybe ask them for this interest. You will be surprised how many founders like especially like first- time found building startups because they just like internally just want to live a lifestyle of a startup founder. You'll be surprised how many of them and you know it looks fancy like just from the outside because you are building some like novel thing with a crazy like like-minded folks like going to conferences raising funds etc etc. Uh, and but it's in reality it's fun until you are waking like in a cold sweat like at 3:00 a. m. because you've had a nightmare that you're running out of cash and you have no money to pay your uh like folks only to find out that it's the reality not a nightmare then all the so and what I'm talking about is the journey of building a startup whatever it's like AI native or not AI native whatever is like a journey of taking like punches and sometimes knockd down uh every time like you were asking yourself this I used to be founder like a few times so every time you're asking yourself like why the hell I'm doing

this like oh uh and if at this moment of time you do not have a strong answer to it chances are you'll just give up but still it's like just a probability like game so you're testing hypothesis so you cannot take anything for granted when you're building a startup so if you have like this strong intrinsic motivation it just helps you to keep going for a long you mentioned about like I think I think it's a negative signal like people should not spend time trying to raise money. people should increase uh uh like MR or or find product market and if you do that then uh the question becomes like how do you choose the best investors because you'll have too many inbounds and the same for founders like you should or same for VCs you should uh like ideally you want to invest in the companies that u you know perform that get traction uh even if it's like maybe before revenue you look at like some you know are they able to hire great people get a lot of GitHub stars are they able to lot to get a lot of attention from the customers etc. So this will change the rules of the like this is not changing the rules of like this will create some um offense defense play uh here where people try to manipulate the other side of the market but in reality it doesn't change anything I think what's more interesting and what's more deeper okay what like what how would how will it look like in five years because on the founder side the amount of opportunities will radically actually decrease increase. Why? Because economy will become a huge reinforcement learning machine that will improve itself and the AI will be kind of you know you will like built an agent and if that agent is doing something helpful for the humanity for world economy it will grow like people will start paying for it and it will get more and probably like large labs have a lot of resources and head start to do it. Soion of it doesn't mean that all startups will disappear but I think the notion of creating a business and the cost will significantly change and also and that also means that the notion of invest investment and the real estate investment will change it. It doesn't immediately um will it won't immediately work with all for all industries but it will progressively capture more and once it works for one like once you have like accounting that is fully automated they will start just like falling one by and at some point it will be significant part of the economy. So it will look like more like public markets rather than uh um — so what do you mean every company will have a kind of a token? How do you mean public market in the preced city state? It means that you can run a large scale it like I don't know who will be doing it if it's going to be like a huge open source network or decentralized network or Google or open AI but just purely technologically you can run a huge reinforcement learning like with thousands of experiments every day like basically have a codeex or clo out like thousand new ideas every day tracking like finding one that like let's say you create like thousand products you put 10 bucks vaccination marketing. You find one that is profitable in that one take. You close everything and then next day you take that one and like generative like um genetic algorithm. You create a thousand more copies of it with different uh you know business models features etc. And this is what Ellen Musk is writing for example on Twitter like that's what that's the idea behind you know XAI and open AI is talking about it. Again it doesn't mean that startups will disappear or venture investors will disappear but I think it will be significant part of the economy and that's interesting how like what you can like I think this analytical and and datadriven approach uh systems like if you can actually build an alpha by this system you can capture some part of this. It reminds me what you said about the hedge funds and how they use right now AI like in every uh tool to invest in public companies. Yeah, I agree with you that there are right now more builders and one solo founders who making the project even without VC uh market. Uh at the same time I kind of from this hedge fund experience because every time every any hedge fund find a very promising deal uh we have a market we have 10 20 other hedge funds make the same algorithm. So basically the market uh themsel push and uh maybe we can also ask Yuri and understand what they think about the future uh because it's really interesting question and how they see this market in five years because investing in AI and it's a very existential one like an honest answer is no one knows actually like AI does not follow like the traditional like this linear like growth path. So that's why uh it will really depend on the some kind of breakthrough moments like they may occur inside like some kind of like top tier like fundamental model companies maybe some kind of a startup but it's really like hard to predict. uh it's like the levels of uh automation like the status of automation. uh for now most of the founders like as stated here the first level uh some of the uh like most advanced are like trying to build some kind of second layer companies but uh like without any breakthroughs I believe vast majority will be like second maybe third layer but still it's like uh needs some kind of progress to maybe fifth layer or maybe fourth layer like some of the companies may achieve not the fifths I do not believe in it like uh in horizon of like 5 years. — Okay, great. Thank you. That's 10. — I'm looking forward time where like complete copy of our uh world will be created just to test one hypothesis. Kind of what uh Stfan said. This would be crazy. Yeah. If there was just a simulation to test just one hypothesis um I think it could be could end up simulation and simulation and all of the simulation just testing u startup hypothesis. — Okay. Uh guys it was very uh interesting topic uh one more time about what we are doing here it AI manset plus ultra VC plus uh cyber fun basically we gathered this event to find out how AI is used right now in real projects uh to find out the best deals for VC market it's rough discussion this uh and we have plenty space for the next time I believe and to find out maybe new flows how it actually used maybe to talk about actual uh points and to make demo. We really love showcases how it really works and we make it every S debate. We talk about founder operation system from different points not only from VC side but it is basically we should cover this one and we covered it from founder side as how they find VC funds how they found new customers or how they operate the mind basically because making a startup is a very stressful uh situation for every founder and you need to keep your mind very clean for this. So one more time. Thank you very much everybody.