Webinar Ellipsis The Zero Employee Ai Native Vc

TITLE: Webinar: Ellipsis - the zero employee AI Native VC CHANNEL: Yariv Adan DATE: 2026-04-19 ---TRANSCRIPT--- talk. Uh so I’ll let Yarif introduce himself and the fund and whatever else he wants to say and then uh

and I’ll introduce myself. Uh yeah, there’s one more. Okay, cool. So I just go. Yeah. Yeah. So um hi everyone. Yeah, I’m Y general partner at at Ellipsies. We’re basically join from your link. Thank you. Two builders that um turned investors. We started in 2023. I spent 20 years at Google building AI products. My partner had a startup sold it to Apple, built some products over there. And we had this dream of building a different VC that is like, you know, super attentive to the founders and doing everything very perfectly. But then we realized that we kind of need to have a hard choice. We create a large fund so we have the resources to to do it. Then we detach oursel to some extent from the founders or do what we wanted to do which is a boutique a small boutique fund where we’re involved in everything but then like you know we have less resources so we basically came with the idea hey can we actually be two people but scale ourselves as if we have a team of 50 people by implementing everything with AI so we basically were AI native from day one which gave us some advantages that we didn’t need to bring AI I into existing processes but we actually designed the fund with the intention from day zero that we want to run it with AI agents. This was like two and a half years ago when at first we were doing a lot of coding and and every time we threw it away and and we written it and basically I want to to talk a little bit you know where are we today and how are we doing it and and the idea was simple. We said like you know let’s really automate everything that we can focus on the things that where humans really need to focus on which is judgment call relationships and and and kind of the strategy and the other thing that we always had in our mind is that once we have AI it’s not just automating but we can actually have a fund that is 24/7 so basically you know have workers around the clock without being afraid of them suing us and and the first insight that that we had was that yeah you know the carim model is quite quite outdated and while you should be focusing like we said on judgment and relations but there is so many operations at everything that you do right there is actually a lot of opportunity to automate in deal sourcing in due diligence in operations in follow-ups all of that was actually a very optimistic view and what we realized is that the places where we need to concentrate to put our brain or brain dump is you know now we call it skills before that it was encoded in other places but one to really have a very detailed investment thesis and we have this like you know currently it’s a giant markdown super detailed that that we constantly update and it’s actually different for we do AI deep techch we do robotics we do metic it’s different in different places but this is basically in a very detailed manner how we think about the world how we think about companies how we do how do we choose them the other place where we made it very important content to make the fund in our image in a sense in in the artifacts that we created. So we defined in great detail how does a team scorecard look like? How does a competitive and market analysis looks like? How does an IC memo look like? You know all of the things this is where we baked our view on how to think about things. And we said okay now let AI operate within that framework. That was the general idea and the two things that we made sure is like hey this shouldn’t be you know anyone that wrote software knows that you have design and then you start actually building the software once the user start using it. So we wanted to make sure that we actually have a feedback mechanism and that a way to keep it live and we made both a passive and an active feedback mechanism. On the passive side um we just allow it to learn from all our interactions from how we review you know what comes out of you know how do we make the decision what we discuss amongst ourselves emails and whatn not. So it passively looks and find mistakes and updates but also we ourselves created the feedback that it’s very easy to come and say oh yeah we know you do that but this is actually not right. So we constantly actually allowed to give it feedback and this kind of detail investment is that now we have 30 different skills that are constantly being rewritten based as we are seeing a real examples of of judging companies. That was a kind of the first and high level and and the second insight was that we can actually automate the entire investment process in what we we thought is a scalable affordable and quality oriented way. Um we automated a lot of other things as well. In this talk I’ll focus mostly on the investment process because I think it’s more interesting. Um and the way we did it we actually broke it into two parts. One part is kind of the sourcing. Okay. I have like a beginning and like you know a beginning of a thread whether that’s a name of a company a name of a founder or a search criteria and I want to decide whether I want to meet these people and prepare for that meeting that’s like kind of one system and then the second system that that we have then is okay once I I am meeting these people how do I go effectively from meeting them to make a decision whether to say no or to invest and and from name to meeting is basically again a very simple process. We start with web enrichment. So basically we start let’s imagine with the URL or whatn not we go to the web and we collect everything that I can I can collect on that or the founder what does the company do you know very simple web enrichment and searches and then we do a very basic and cheap basic fit search right so we start with a very large funnel we just want to see does the stage fit that at a very high level does this alignment fit so we do like you know kind of very crude test on team on stage on and on TZ’s feet And that already throws a lot of stuff. If we get a positive answer, yeah, this is interesting. We actually can automate the outreach to the founder because we have the some of the contacts and we say, “Hey, we are ellipses. We invested that. We looked at what you’re doing. Looks very interesting. Would love to get in contact with you. Can you send us a deck?” Right? So now we got a deck. Now we do a slightly deeper analysis on the deck. And now this is domain specific assessment. If it’s robotics, we look more at the um or if it’s the deep tech, we look at the different criteria and the output of this is like should we meet them? Yes and no. And what what confidence do we have in that answer? And everything so far is completely automated and like you know takes minutes. Um, if we have a clear yes, then we actually run a a a premitting analysis, which is a deeper an analysis based on the deck. And I’ll show how that how that looks like. And then we email the founders, hey, let’s schedule a meeting. And often we can actually already send them a bunch of questions that came with the analysis. If it’s a no, we can automatically say, hey, sorry, don’t want to meet you. You have a great company, but here’s why we don’t want to meet. And of course we can if it’s in the anywhere in the middle we always can escalate to a GP and anyhow there is always human looking at that. So again very simple straightforward automation process with very clear ways of quality to hill climb and improve and to monitor and to eval. And I can show you you know a little bit the tool that we built. We look at the at existing tools but we we figured out that you know like any SAS tool you pay for something you use only 5% and it does only 40% of what you want. we are like super super early stage. So like most of the tools are not really for us. Um so yeah so this is like me clicking here and this is the the real tool. So this is the tool that um scrapes everything and we choose our own sources and what not right? So now there’s like I don’t know 40,000 here and basically based on the initial right so we see here what does the company do? We see like based on this very initial criteria, you know, TESS feed, team signal, program, you know, traction data. We can see the sources. We can see, you know, kind of okay, who are the last companies that the top ones in terms of scoring based on that. So, we can look at the the ones that are more interesting for us. You can always look like you know who are the most recent ones and you can play here with a pipeline and if you like something you can click on the quick screen and that generates you know to to the next level and of course a lot of that can be automated. Um so and the cool thing is that you know we constantly update oh we don’t like this we want to add new criteria we want to add a new button called the whatever usually it’s a few minutes of change and because all of these skills of how do we look at the company how do we connect and whatnot we implemented now from the beginning it was very compartmentalized so you know it was separate but now it’s implemented as a as a cloud skill we can actually call it from anywhere we And anyone that implemented here anything with skills know that skills are not very easy to work across because they sit locally in a folder. So what we did we created an MCP that points at the skills. So anything that has an MCP actually knows about all our skills. So that means that we can actually access that skill from any place. So for example I get this list of companies from plug and play. What I can do is I can take it I can just throw it to cloud on my phone or anywhere and because it has MC I can just say use your skills to extract the startup data and check relevancy check for ellipses that’s it you know and that’s enough for it and I think someone said that claude might be down but yeah so the result that I would get from claude it’s like yeah okay cool okay this is the relevancy check there are 36 startups two are relevant four potentially irrelevant 30 are not irrelevant and And it will go and explain why T mastery TZ is fit mode type why now right this all goes to the very detailed criteria that we put inside it and and yeah you know potentially relevant and so forth right so this like actually makes our system you don’t need to go into the dashboard and tool we created and then you can also of course converse with the model because it has access to it I can ask questions I can ask change it what about you okay you misunderstood right so it’s like it supports the whole shab you get the power of the model with the knowledge that embedded in it and I mentioned before um that we have also like should we invite to first meeting right so this is a slightly a more deeper analysis so we get the decision you know this is from oppus it says a conditional meeting confidence medium it does the analysis that you know top three chances top three concerns below it actually it’s too short but below it also has here are the questions that you should send to um to the founders and one question we had okay how do we make sure that at this stage we don’t make mistakes it doesn’t hallucinate so we came with a very simple solution because it’s such a a relatively cheap analysis we actually send it to all three models so we said to GPT we said to Gemini we said to cloud um and we do a lot we did a lot of evals and we found like you know funny things um so the first thing we wanted to make sure that the thing doesn’t make up. So we run like you know 15 times on the same model, the same deal. We saw okay it’s very consistent. But then we saw that out of 15 deals um Gemini said 14 yes one no chip said one yes 14 no and Claude said five. And when you look they actually agree on the fact but like it’s a little bit of the character of the model. one is like you know I think Gemini is an optimist you know this guy is a pessimist and this one is more balanced and and like a manager you know your first thing oh you know should I regularize that or whatever and we thought no actually this is great we have a team now each team member comes actually apparently with some bias let’s use that so you know the query on top says hey just so you know this guy is a person usually you know take this with a grain of salt or tell us why he’s excited and we actually created um a this chat experience where we we get when we get the result we can actually you know talk and get you know model differences what are the key concerns why this decision and have a whole conversation with the models to dig a bit deeper into these decisions. So, so that’s how we did kind of okay, how do you go very quickly in a way that you can control quality and scale and bake your own thinking on the first stage and then we said okay cool now we we get to to meetings you know how do we automate the next level and here we realized we should think about it as an as a loop and the loop is always we we have information so far we have whatever we know about the company then we have gaps red flags and key questions now we take is and now we go and we talk to we go and collect information, talk to the team, talk to references, do our own research to answer these gaps, red flags and key questions. Now run again, see can we reach a decision, a yes, a no or still gaps and just loop like that until we we get into a a clear decision. Again, very easy algorithm to go implement. And the way we went on to do it is like the mo the moment we get the deck, we already create the full IC scorecard, IC memo, team scorecard, we have everything from the beginning. Just at the beginning there is of course more gaps and more questions and then like we are just working in this loop to fill it up to update it and every time we meet with them or collect we transcribe everything. We have an agent like I said every day looks over the emails find from all the communication and everything that we had find any information that was shared with us and again pushes it into the deal folder and runs again the analysis and we made the system intentionally stateless. So there is no state nothing in the system it’s only in the deal folder. So basically the mo so you can change the model change whatever you want because all the information in the deal folder just eats it and then it does whatever you do it right and you we have versioning so you it can actually sees what evolved but like because it’s stateless actually it’s a little bit like what karpathy recently um pointed out so yeah so just um and then we have like I said you know evolving scorecard like I said you know from email conversation meeting notes sl IC discussions data room any information that comes live or whatn not we push push it in and it just realizes that it needs to run some stuff. I can show you a little bit how it looks like. So that’s our system. It’s called Olympia. Now you see all our deals here. Yes. Yeah. So yeah. So this is a deal and of course this is a deck that we we took from the internet. It’s one not one of our things but like Yeah. What what you see here is um so this is what we get the moment we upload the deck right so that that’s already cool right from that point we get kind of the score based on the very detailed analysis tells us where it’s submitted this is kind of the summary it gives you know the team team verdict a little bit you know some deep deep dive some highlights you can go here and of course you know see the actual deck you know that’s the you can see the scorecard you know, funding teams, all of these things, deal breakers, PMF gates, these are all sorts of flags and criterias that that we defined in how we want to see things. Like I said, red flags, top strenges, top concerns and you can see always there is the full report that you can actually go into the details and everything as well as the sources and references. And every time there is a fact we always and I will touch a little bit afterwards how we deal with hallucinations and and and founder But any fact we have on you know is this verified is this plausible or is this like kind of seems and then like you know yeah we have the deal memo already written from the first momentum. Yeah that’s the enrichment team assessment. So you can actually see you know the score of the team quality you know how do they go on the different dimensions recommendation key risk right so very detailed um section for each one of these that that continue to build and if we don’t like something we give feedback it changes so it’s all in our full control me and my partner and the main coders here and then we realized that hey it’s very cool that we scaled ourselves the two smartest people on the planet but wouldn’t it be nice to actually get the opinion of less smart people as well when we see a deal. So what we did we created this kind of simulated investment committee. So we took a bunch of like super top well-known investors from the valley and others that invest in super early stage deep tech and actually you can scrape their profile and create the copy of them that I’m sure is better than how they represent themselves. There’s so much information about them. So, we created a profile of a a quite a few of these investors. I hid a little bit here. Here are the names, but yesterday someone asked, do they know? And I said, okay, this is going to be recorded. I’m not going to have a screenshot of their names here. Um, but so, and basically they are having an investment committee. So, every time we choose a bunch of them, run at random, we choose who’s leading the IC committee. He starts he kicks off the meeting. Everyone has a statement then they have some discussion and then they have a summary recommendation. So basically here you can see you know they’re saying the recommendation we should pass. You can see you know consensus what are the strengths you know unresolved tensions and recommended for next step right send clean friendly pass to right they literally tell us what to do. So basically we are borrowing the the smartness of of other investors as well and we added to the mix and and of course there is a lot of new improvements here and and of course I’m going to create an Alec Shamis version as well after he explains how he makes decisions. Um yeah and the cool thing is that like you know we have like I said there is always intelligence that is running behind. We constantly, you know, 24/7 we monitor um email, calendar, uh LinkedIn, the the internet, right? So, so you have like um a bunch of agents running around and collecting information and generating information. And the cool thing is that once we created this stable state and at this point we are like in in a decent condition. And I’m sure that like you know it will continue to improve. You can start building stuff on top of that because once you have all the information why do you like what you don’t like you can start writing ops agents to make sure that you you’re not waiting too long that you’re not dropping stuff off the ball that meetings go. You can run marketing stuff that actually shares with investors and partners here are the things that we’re doing or here’s what our companies are doing. Here’s what we’ve seen. Here’s some insights from what we’re seeing. You can use it investment creation. Right? like actually value compounds once you have the daytime system in place in a in a stable situation very quickly how it’s built like I said you know um we constantly evolved at the moment basically we have a virtual machine in the in the cloud running on the on GCP um all the the deal folders and stuff is just a Google drive with folders per per deal we use NN to for schedule stuff that we want to make sure that happens you know properly and then we have an MCP server that is the like I said the entry point and we have now like I don’t know like 40 different skills and every time it spans a cloud code a a session with a skill so that runs and and have access and then we created all of these we have open claw as an as our chief of staff we have the deal chat app that looks at that we have the sourcing app we have automations so we add as many clients that we want and because it’s MCP based it’s like you know today you can connect almost everything to that it’s less than 500 a month I think we’re under spending I think we should actually find ways to spend more yeah I wanted to to talk a little bit about insights and stuff so one one question is like um how do you deal with hallucination and like I said there are two problems one is how do you deal with hallucinations and the second one is how do you deal with like you know because pitches are pitches you shouldn’t believe everything and Our insight on hallucination was that we cannot stop the models you know that they hallucinate by nature but we need to make sure that these hallucinates don’t lead to bad results or bad recommendations. That was like kind of a technical assumption. Um one is like I said you know in the critical places we actually run multiple models where we’re like we’re actually it’s like evil on the facts. So we find that and we we do the evil. So so that’s good. The other thing which is very easy is that models are hallucinating when they’re forcing to. If you explicitly tell them, hey, tell I’m not found, I don’t know, or you know, gap or whatever, they actually do that instead of of of inventing, that actually solves a lot of problems. The other thing, like I said, that you need to classify everything as verified, claimed, or suspect with a source. That also actually reduces all of the stuff and allows you to check where it where it’s coming from. The other thing is always like find stuff from multiple direction. If we’re looking for founders, we’re looking for company, don’t look for a single from a single source and a single signal. Look for multiple ones and then dupe them. Um the other thing is there is many more hallucination in all the human readable crap like you know the actual outputs of reports but actually a lot of the work and the stuff that we move in between is JSON with most succinct information. So even if there is some hallucination or like you know inaccuracies the reports it doesn’t really matter because the actual information that is being passed and proved is is much more succinct and we actually do some redundant work where certain skills could use like the output of another skill but we actually intentionally run and check for inconsistency. Again it’s cheap and it’s a great way to to make sure that we don’t have hallucination. So these are some you know very kind of easy to to do and and the cool thing again is when we see an elucination we everything is logged so we create a lot of logs and trails and we actually tell the model hey dude we see this hallucination please find out why let’s find and fix it that it doesn’t happen again usually actually comes up with we solve it the other thing is how do we train to have critical thinking like don’t believe what you read don’t believe what you are told so one thing is like if you just tell the model um look at things to make reasonable or don’t believe you don’t get good things. So what we actually did we we gave like a a pretty long table of facts that don’t make sense like what is like you know a healthy LTV CA you know so we actually went and and gave a lot of things and said okay this is reasonable this is unreasonable and then they’re much better at finding things and not just like you know making stuff up you know then we have like you know does the does the traction make the stage you know does the deck contradict our customers you know are the naming concrete entities so we we give it some some things we make sure that the scores are calibrated. We also have this hard gate I showed before like PMF or whatnot. So even if the model for whatever reason decides to give the wrong grade, we still have like additional gating factors whether it’s a like you know positive like outlier team or big opportunity or whether it’s negative on certain things. So we um we don’t allow a hallucination or inaccuracy to actually drive a decision and like I said example is five simulated investors who disagree and so everywhere where we can encourage disagreement critical reviews data that contradicts actually you know it actually forces and grounds and we constantly add it based on real examples before I finish I’ll share some lessons that um that we learned that are good practices. I mentioned some of the um one like you know really completely stateless skills the deal folder it it simplifies the system and allows you to constantly optimize it without worrying about you know backwards compatibility or whatn not. MCP is the universal trigger. Anyone that worked with skills know that like okay sometime in the next few months they will fix it but at the moment cloud desktop is one skills that is like you know stored in their folder. Another one is like cloud code is a different set of skills on the phone you cannot use it. So we created that entry point with MCP super simple super efficient and it’s available from everywhere. Um three you can vibe code a lot of stuff but actually logs blockers you know dduping hashes for files and whatnot you need it doesn’t do it. So I think like some knowledge as an engineer of what a good system is on on some of the basic stuff you need to have someone that that knows how to do that to get a good system and really run hygiene. So we run on a daily and weekly basis like hey find security problems or where are all our tokens going and you know one time we run it and we found out that we are running this crazy deep research report and half of the tokens is actually going into once the report is generated just to create a doc x version. I don’t know why we even created that but like it actually was more than creating the entire report. So we just drop that and have an MD and MD view boom half the time half the token. So like actually running and making sure what is what is causing latency, what is causing the rip and they are very good at actually if you put the right logs they’re very good at knowing it. So really some simple tips. Um we asked ourselves why do we need humans at all and we came with four answers. Some of them are already obsolete and one we are super early stage and you know sourcing is great but if we source it others can source it as well. You definitely want people in places where you know where companies are not out yet. So we exist to actually find people before they are sourcable. The other thing which is obsolete there is a lot of and lies in our industry and we realize that like you know and lies don’t survive six hours on face to face and drinks. But then I realized that I can actually put a an agent in a in a company for a week and listen to everything that is going on in the email in the documents in all the meetings have it completely local. It doesn’t send anything out just at the end it will tell me or not and I think that that’s actually a viable solution. Another thing is left field questions. I think the agents is much better than us in doing the research, analyzing it and coming up with all the smart questions. It is definitely better than us. But many times in a meeting you ask kind of a left field question and that kind of reveals I cannot trust AI at this point to tell it hey from time to time ask a left field question it will be very left. So, so for that we still need people and we still believe and maybe it’s wrong but we still believe that like you know some of our charm human and wit help us win some of these deals although Ellie our open claw chief of staff I think is actually charming more charming than us at this stage but yeah but but these are like I think like the main places so again relationships are some judgment call where humans are required I’ll skip that and yeah and basically Yeah, you know, AI gave us time. We spend it with the founders and of course with creating more AI. Shout out like, you know, it’s mostly me and Matias, but we have some young people on the team, students that it’s so cool to see because, you know, they didn’t know how to vip code, but but it’s so powerful today. So, Serena and Louis on our team actually helped us build some of these systems. And of course, Cloud Code is the number one engineer on the team, contributing millions of lines of code and never wines when we throw it away. And yeah, that’s it. And I hope I’m on time and can hand the stage to Alex.