Building an agentic hedge fund

Hey guys, hope you're well. Uh today going to walk you through an AI investment finder uh that I've built on Reddit. I have spoken about this before in a previous video, but I've now made like quite a lot of changes and improvements to it. Uh I've run about 700 different trading algorithms uh that I've been developing. Um everything completely like vibe coded. I've got returns from 180% win rates at like 90% by by win rates.

Sorry, just to clarify, I mean profitable trades. So it means in like 90% of the trades it's profitable. I've got loads of them at like 85 80% uh and and 70% which like is kind of unheard of. Um, and yeah, I think that this is uh a a kind of new and expanding use case for AI that I I think will really explode. Uh, one kind of recurring theme in my life is that I generally tend to be quite early to things. Um, I've seen that with like Bitcoin, uh, with like the first Bitcoin I bought was I think it was about £80.

It would be like $120. I was very early into like no code stuff building like these automations like eight eight years ago um with the OF side of things as well. Just throughout my career I tend to be be very early to things LinkedIn automation. Um and I just I have a feeling this is is a space that is re really going to explode over the next few years. So like AI, investment finders and and helpers. Um and yeah, I was very well positioned uh in order to build this and it was using all of the skills um and experience that I acquired from building the trend finders that I was using for finding viral content for OF models on Tik Tok and Instagram.

Uh where we would scrape hu huge amounts of data, hundreds of thousands of profiles to look for the most viral content. um and and what was actually causing that content to go viral. So, it's the exact same skill set just applied onto at this current stage onto Reddit and looking for like what are sentiment and trends within retail investment and how can we um how can we exploit that to to to make a profit basically. Um ju in terms of like other platforms as well just whilst on the subject of um Instagram and Tik Tok I am actually looking at like currently thinking about expanding into them because we are in a position to be able to scrape pretty much all data um and by all data I mean like all users of um theoretically of both platforms um and you could look for uh for for ways of utilizing that data to to to make investments but just by checking things like comments, spikes in interest in certain areas, what people are talking about, specific products or even companies that start to get mentioned a lot and and you could come up with like a different investment theses uh around that.

Um but anyway, just to go in into into the system, um this is like the air table and kind of like the the structure of the air table. So on the far left um side, this has all of the stocks and all of the mentions. Um and I'm just pulling in the comments, the sentiment score, how the sentiment score changes and various different things. Um number of comments that that they're they're receiving. Uh average number of up votes, the last time it was mentioned, uh the number of unique commenters I also find to be like a very interesting one as well.

Uh the current price of the stock, this gets updated every day, the market cap of the stock. um total number of mentions, uh price increases, total number of up votes over different periods. Um velocity of mentions and how that how that increases. Uh the first time the stock was mentioned is also an interesting one. Uh return since first mention. Um yeah, quite quite a lot of a lot of fields going on here. And then at the end I I'm using like the Polygon API uh to pull in things like the logo and and the market cap and the address and also the sector uh as well as like the number of shares outstanding.

Um so like the kind of the the technicals u related to to the stock itself. Then in here we have um all of the actual comments. So this might take a minute to load. Yeah, the these are like the the raw comments. So uh we have the author of the comment, the stock itself that the ticker that is being mentioned, score is the karma, title of the thread, um any replies that it makes to the comment, the price at the time of the comment, uh the price 24 hours later, um and and the price 7 days after the comment.

So we could see how accurate they were at predicting kind of price movements over the short term versus o over the long term. Then we have comment stock analysis uh which I I'm not not going to go into but that is how I do the AI analysis of every single comment um that gets brought in here and we determine the kind of sentiment and and uh the the trend behind it. Then commenters actually shows you basically the every single unique author that has ever left a comment.

Um, and it has the number of comments that they've made. It also shows you the comments that they're listing. Uh, like this guy is like only talking about Reddit, so he's not that diversified. He's kind of just spamming um one one specific ticker. Uh, but what you can actually work out from this data is how accurately each different author here is at predicting good stocks to pick. Now, just to save you a bit of time, I can say there isn't one person here that is actually incredibly good at picking stocks.

What the way that the system works because I I can can literally see the data and I can tell you that hands down. Even looking at things like the total karma of a user, the average karma, how long they've been on Reddit, these aren't major indicators in terms of um how good they are at at picking stocks. What is a very good indicator is when you find that the subreddit as a whole takes consensus behind the stock. That is a very good sign.

But what you really want to be doing is figuring out how do you get in early when there's a couple of people talking about the stock and there's maybe steady and consistent interest around the stock, but you want to get in before everyone starts talking about it. And I've seen a couple of other like Reddit, I wouldn't even call them trend finders, but they're kind of just like Reddit trackers and they track the number of mentions of specific stocks over a specific day.

And there's even subreddits where where they just post in there every day. In my opinion, they're doing it completely wrong. I mean, not even in my opinion, this is what the data literally says. They are doing things like when interest in a specific stock spikes, for example, if a hundred unique people start talking about a stock in a day or even hundreds of unique people or thousands, they're considering that to be a bullish sign.

And it's not. It is it's actually a really accurate predictor of the stock is currently at the top. And I can show you that in a couple of like charts and graphs um that that I built out as well. So again, I'm just kind of explaining to you a little bit about how the system works. And you do not want to be jumping into a stock when the whole of Reddit is jumping into it. It is it is one of the worst things that you can do.

And I I'll show you that in in um in just a moment. Just to clarify as well why I'm sharing this because like you know I'm giving you like information. I think this is like a billion-dollar strategy by the way and we'll go into more into that in a second as well. First off, I'm not kind of giving you everything in terms of of how how it's being done. I'm just giving you some of like my higher level learnings in terms of what really works in terms of investments, but it's kind of common sense to be completely honest.

It should be a a common sense thing that if every if there's massive spikes in interest in a specific stock, it means that basically it's seen a large price movement and Reddit is getting very excited and everyone's jumping onto it and that is not a time that you want to be in that stock. Whereas all these other like Reddit kind of again they're not trend finders. They're just kind of tracking interest in a stock. They're they're they're just do doing it completely wrong.

But why I'm sharing this? Number one, I'm um and sorry I should just clarify as well. These are all back tests. So today is the first day that I'm going to be investing capital into this system. The only reason that I held off is because I knew that there were a lot of there was like just some low-hanging fruit, just quick wins that I could find and I've just now found them. I mean, I knew what they I knew what I wanted to do.

It was just a case of like building an algorithm in a way that was able to actually detect it. For example, detecting peaks in in um in interest and selling at at the height of the market. And I I've now been able to to figure that out. So now I feel that I'm there's still a lot of improvements and a lot of ideas and directions I'm going to be um continue building, but I'm now feeling very very confident about putting my own money into this.

So I'm going to be putting in $100,000 um across uh across four different strategies. Um but why I'm sharing this? Uh I'm for one considering opening like a hedge fund/investment fund. And I'm just kind of on one part interested in like feedback but also in like potential interest uh from you know there are some very wealthy people who who follow my channel. Like I just spoke to a a gentleman who was in my coaching before who does like 3 million a month through OF through his agency. uh and and he was interested uh and asking like questions about the like investment stuff and and potentially doing something in um if I was to open up a hedge fund.

Um but so that's one thing is like I am considering open up a hedge fund. Like I'm also very interested at the moment in businesses. Um, in I'm only interested in companies that are extremely high profit per employee, seeing as I'm the only person working on this, uh, and have built absolutely everything aside from a another developer that I I work with on other stuff on OF who built out something for like the Sharpe ratios like the actual strategies and everything itself um, is all like developed and all of the ideas have come from me and this is a business that has the potential to become like a multi-billion dollar um, company.

Um the other like thing as well is in terms of other businesses that have the potential to make m multiple billion um a multi-billion dollar company at the moment really in my opinion is SAS and I'm not really sure what else. And a lot of the SAS that we're seeing at the moment, it's kind of just wrappers around AI, um, you know, like a building like some type of front end or some type of service that is just plugged into like GPT or into like whatever model on on the back end is kind of what what we're seeing with Cursor.

It's what we're seeing with um, Claude and just just and I've got nothing against that by the way. Um, you know, I have like complete respect for the the entrepreneurs behind it. My biggest concern is that I think the value that you deliver can kind of vanish overnight. For example, with Cursor, we just saw uh Google release their um their alternative to Cursor and Cursor is also just like a fork of um of VS Code, which is is open source.

Anyway, so again, I do think that they are doing some some like really interesting things, but what I mean is like with the advancements of AI, we can just see value kind of just get eaten up overnight. And I don't think we've ever ever actually been in a um in history being in a position where value can be created but also be taken away so quickly because of the the advancements in AI. So basically I'm just very kind of worried about if I was to build a SAS and spend like a year or two years working on something and it just becomes completely redundant in two years which is a very high possibility of that happening.

Uh it's just not the most um enticing business model uh to me at the moment. Uh, another reason that I'm very interested in like the hedge fund side of things is I was watching a documentary on a guy named Steven A. Cohen. Um, this is a bit of like a controversial example to give because from my understanding he was known for um, insider trading but he amassed a fortune of about 23 billion and it it was due to insider or or allegedly supposedly I'm not like super up to date on the story.

It was a documentary that that I watched, but um he was able to generate 60% returns per year, which was just considered to be absolutely insane. And he was um able to charge a 50% fee and a 3% management fee. So 50% of profit is what he would take from his clients and a 3% management fee of whatever capital he was managing for those clients. And that is how he was able to to accrue a fortune. And traditionally in hedge funds, the fee would be a 20% fee of um of of profit and a 2% management fee.

So his was like extortionate compared to kind of like the industry benchmarks. The reason he was able to get away for it is because get away with it is because his returns were so high that people were were willing to pay um to to pay incredibly large fees in order to have access to to his his funds. Whereas my returns, you know, I haven't got the the history which is is the only thing really going against me are substantially higher than uh than anything in in that range.

And again, not just on like one or two strategies. I have got hundreds of variations of strategies that I I'll I'll show you in in just a second. Um the other thing is I think that like our expectations I mean not even my expectations but particularly like the younger generation their expectations from um just from what they've seen in like crypto or meme coins of you know people making like a,000% or 10,000% or turning like a couple of hundred into a few hundred,000 are just completely like blown blown out of proportion.

And just again to give you like an example of this, uh, in my like best strategy, let's say I started off with $100,000. My best strategy has about a 180% annualized return. So if I'm to annualize this o over 10 years, let's say, you would turn $100,000 into three billion dollars. Now of course if we look on the graph like for the first few years it seems kind of quiet like you know just by year five I mean by year four you've turned 100k into 6 million which is of course insane returns but towards these later years is where it starts to get insane.

So my priority basically and uh there's no guarantee that I can kind of can kind of do this by the way I expect this to work at least for like a couple of years based on my my back tests and and historical data. Um but really the challenge needs to be of how do you maintain this level of returns over um o over an an extended period and even like improve on these returns but even on like a more kind of reasonable rate like I've got hundreds of strategies at like 80% yearly returns o which is much more reasonable in terms of being able to maintain over 10 years you would still turn $100,000 into $35.7 million um and even over like 20 years you know I'm over 30 I'm 30 years old now.

So by the time I was 50, you you would be sitting on about 12 billion again under the assumption that you can kind of maintain these returns. Of course, there's going to be like fluctuation and so on, but this is just to kind of give you an example of like to become insanely rich. It's not that you need to make, you know, be gambling on like meme coins and make a,000% in a year uh or or something like that. really you want something that is more in my opinion anyway more systemsbased um that is kind of proven and and logical.

So that is on one level I I would just be interested in how much interest there is behind opening a hedge fund um because it requires like upfront capital and a lot of like kind of legal stuff in order to actually open it. Um and if I'm able to like prove not if just it needs time for me to like prove these back tests over a period. uh if I can sustain these returns, I would just have people like queuing up to to to to give give me money.

Um on another sense, like the the other way of kind of monetizing this that to be honest, I'm kind of against, but I do keep thinking about it. Like it does make logical sense is selling the trades as signals, which the reason I'm against it is because it's just known for being like a very very scummy um business. Um, and the other thing that does come that people would ask is like if the signals are so good, why are you not just using them yourself?

And I do plan on just using them myself. The reason the only reason to kind of sell the signals is the only thing like limiting my returns would basically be like upfront capital. So what I was considering doing is having the AI build something. I don't know if people are familiar with this, but it's called Alpha Arena, which was an experiment where a guy g or I I think it was actually a hedge fund. They gave all of these AI models $10,000 each.

Um, and currently the only two that are profitable are Grock 4 and GPT5. Everything else has lost money. Um, but you can actually click in and see all of the transactions and and everything that it's doing and and so on. So, I was actually thinking about doing something like that, uh, where I would set up maybe like four or five of my my different, um, strategies and just kind of show them show them up here so that you can like, uh, see everything, see the actual trades that they're doing.

I would do like a delay on displaying the trades of maybe like 30 days. uh unless like someone wanted to pay to access like the real-time data and you know you could send those signals to like an email or to telegram or to whatever and then any business any money that was made from monetizing this like I have no need for um for like dayto-day money there's nothing that I kind kind of want to buy what I would do is any money that is raised through this business um would literally be given would be invested straight back into the strategy as to increase the the rate of compounding.

Um, so that that was like again it's just me kind of like brainstorming and and thinking out loud. Um, another thing to say is like I'm not the only person that's uh that's exploring this. In fact, this guy has been looking at like social arbitrage for the past 20 years and he's turned something like $10,000 into 80 million uh looking at things like Tik Tok comments. His strategy is quite different to to mine, but again, it's um it is it's a 3-hour long interview.

You can definitely watch this on 2x speed. In my opinion, that there's quite a lot of waffle um but I'm also quite quite guilty of that um as well. And then finally, the other reason is uh to to give this away um just like this kind of opportunity is because there's I don't think that there's a lot of people who who would be capable of of building out something like this. But again, I I may be kind of proven wrong particularly as AI starts to get smarter.

But if you are working on something similar to this and are seeing like similar returns, uh I'll just go through like some of the strategies as well. Basically, I would be very interested to um just to like network, share ideas uh on what what we're seeing as as working. And the other reason that I'm not too worried about the information be um being out there is that uh it it doesn't actually um hinder me if other people are making these like trades at all.

But anyway, so here are like a load of the different trades. There's currently 455 here. I have done about 700 but I actually had to delete them as I was hitting the uh air table limits. So we've got Sharpe ratio being like calculated here just to give you some idea like um [snorts] Sharpe ratio is kind of a measure of like risk and reward. uh and a good Sharpe ratio. 0.5 is like borderline. Anything above one is good. 1.5 is very good.

Two is excellent. And anything above three is extremely rare. So here we have 1.5, 1.5, 1.5, 1.2, 1.7, 2.5. Really interesting strategy. The reason this is so high is because it has a really good win rate. So the returns are at 164% and the win rate is at 83%. This, by the way, this is all dynamic and it updates automatically. Uh, it's based off of of rollups from the comment daily stock price that I mentioned as well. Um, so yeah, and here we have the buy count.

So, this shows you the number of trades that were executed. So, of course, for like statistical significance and for like a context score, generally I prefer strategies that have a higher number of buys and it means that the win rate is actually more more consistent. here 86% win rate really nice and again just so like the top performing strategies are like above 100% returns and I did have a load of other variations of these but they were kind of very similar so these are like the the foundational ones that I decided to keep um and so how sorry how many strategies have we got here 23 strategies above 100% but then if we kind of scroll down here in terms of strategies that are above like a 40% return which again in the business of hedge funds and I didn't actually realize this.

It was only when I had a friend show me who had access to like a Bloomberg terminal uh where they can actually see the returns of um of other like top performing hedge funds. And for the past year, these like, you know, very smart people on Wall Street in fancy offices making exorbitant amounts of money, they were ranging like 16 to 22% returns. And so, anything like I just kind of tend to delete anything that is below a 20% return.

Um, and I I've got um yeah, 315 strategies that are above a 40% return. also with um with really good win rates. So then just to go a little bit more into what I was talking about as well, every single strategy, this is essentially a back testing um engine has its own chart. So you can actually click into this running on a different port. You can click into this and it will open something else that I built which is a a charting tool.

So, I'll just like clear this up for you a little bit to make it a bit more clean. So, the blue lines basically show the mentioned daily number of comments and you can see where it kind of spikes up and then the uh gray line shows the price and then if you see where it says normal or it says spike that was something similar to like the fear and greed index that I I I was working on building myself. So it tries to work out um the kind of fear and greed and like the sentiment structure of of Reddit um as a whole.

And basically what you want to be doing the green markers by the way they symbol the buys and these are a load of like technicals. So I was building out various things like histograms um the RSI index uh that you can can put on top of there as well. um standard moving averages, RSI of of uh sorry there there there's the RSI. So you can kind of track all of these different uh things like like Fibonacci levels uh the volume we we can put on there as well.

Um and you can kind of do comparisons to come up with like a a visual looking for patterns and and I have also been using by the way like um machine learning uh is my first exposure to to using machine learning actually. So giving like large sets of data to try to look for patterns and then also using AI to do image recognition on these um on on these charts to look for patterns as well. So the AI is actually able to um to use to go and analyze the charts and look for patterns and comes up with like a thesis and a hypothesis on what would be a good back test or a good variation of this back test to run based on on what the charts are telling us.

Um but kind of as I was uh was was mentioning what you really want to avoid is the spikes that you know uh we didn't know that it was going to continue rising but that spike continue uh signaled a peak. So prior to this there had been little chatter and then when it peaked up here that was actually a spike on the price and then of course it kind of cooled off a bit. this is a bit more of a consistent range and the algorithm made the buys here in this like accumulation zone because it was like a healthy range and then here we're seeing spikes and it stopped um and we've profited 102% with a 100% win rate across 13 trades.

The reason that you don't see 13 uh triangles here is because these are based off of weeks. So it could have it will have performed multiple buys uh within within a particular week. Um, and so that that's that's the reason for that. And then we can just flip flick through them. I can look at all of the strategies here, by the way. I can just open them up and it will load up the dashboard and I can just flick through with my arrows.

And this is showing the ticker and each of the buys uh that were part of that particular strategy. Uh, so again here, this is kind of like an unhealthy range because it's like spiking up up and down. So we want to be regardless of the price we want to be buying when there's like steady and consistent chatter and again a 70% return 100% win rate uh across these uh trades as well. This is a really healthy healthy chart. Uh so again buying at steady and consistent chatter.

The reason it didn't buy at the start here is because on this particular strategy it looks for um it doesn't just want to buy immediately. it kind of waits to have a bit of trust uh around there being like a consistent level of interest before initiating the uh the buys. Um this doesn't have any sells on it as well, but I can open up a strategy that has uh it does have sells actually. Um I think they're just not being plotted on the graph right now.

Um but yeah, you want to be accumulating in in these ranges. And then you can see my um what I built of the uh spike ratio. When you see a spike in comments, you can see here it says spike. That is something I mentioned like the fear and greed index. So in the um algorithm, we could literally just say sell whenever a specific stock enters a spike zone um offload all of the assets and then you could start to look for it to cool off and to accumulate uh again.

So, uh, yeah, this was something quite interesting to build as well. And then you can click here and it will download all of the screenshots, um, of all of the, uh, sorry, this will download everything. This will download just this one chart. This will download everything for all of the charts, for all of the buys in this strategy with various different combinations of these and it will give it to an AI to analyze uh to try to to work out what what were the best um what are like some of the uh patterns that that are appearing commonly in the charts.

Um, then another thing as I mentioned like this was all vibe coded. So I just think it's worth me like showing you a little bit about like my current setup for for vibe coding. So what I my my current structure for for building out all of this stuff is um uh on this left hand side I like to have a project management agent. Currently the project management agent is uh Claude Opus 4.5 but it kind of changes all the time like GPT 5.2 was released yesterday and I was just playing around with that this morning yesterday.

So we can call this like a project management/brainstorming agent. The reason that I set things up in this way is I find that you get much better answers, you get much better results, but also to actually preserve the um the context window of the agents I find to be one of the uh biggest factors in terms of getting better results from your um from from your agents. So this is like a project management/brainstorming agent and basically I work off of GitHub issues which it's just like a task management thing.

But the reason that I like it is because you can just press here and it it will create create a GitHub issue. But when I say brainstorming the other thing that I tell it to do is to ask me questions to help me shape and to simplify ideas. Um so it will say like oh what about if we do this? It can it can suggest like an MVP as as an alternative. Uh and everything will just be kind of made from here. And you can see here like it asks me um asks me like different questions.

Then >> your agent needs your input. >> On the right hand side here I have um these are what I call the dev agents. So these are three different agents. And so what I would do is once the task has been made here, I can say um can can you work on GitHub issue 55 or whatever the num the task number is that this agent has just made and this agent will then go and work on it. And I've got two other agents here. These are all currently Claude Opus 4.5 which is currently my favorite model for programming.

Um but previously it was uh Gemini uh sorry uh GPT5. Um, and then I just kind of felt GPT5 actually went downhill. But currently everything is on Claude Opus 4.5. Um, which I I find to be be the absolute best programming model. So these three can work in parallel and as long as they're not working on like the same area of the application. Um, as long as they're not trying to like edit the same files, it's completely fine. So you can have three different kind of like developers working at the same time.

And then on the far right here, these three are what I would call like testing agents/QA. So quality assurance. So when this agent says that GitHub issue 55 is done, I'll say over here to this agent um can you uh like the dev agent issue 55 is completed. Can you please review it? And the reason that I do this is because if you ask this agent, it's going to say, "Yeah, I've done it. Everything is completed." You can ask it to double check, but it's kind of already got in its context that it it's completed.

Whereas you're coming over to this agent that has a fresh context window. Now, there's two things that that um like or an alternative way of doing it. Instead of using Claude Opus 4.5 again, I could actually use a different model here. So I could choose to do, you know, GPT 5.2 or Gemini 3 Pro, whatever I want to do. Um, I can have a different agent go and check that same issue and it will check the requirements, what I'm expecting to be, um, what's the acceptance criteria and everything like that.

And now a different model is checking the work of Opus 4.5 to verify that it's been completed and match my requirements. Also, when I create a task and an issue, I generally like to give context of my expected behavior. What is the actual problem that I'm trying to solve in this particular issue? And I find when you give kind of like a user story, when you give the agent that context of what is expected to happen, it helps the agents to make much better decisions and to actually understand what what um what it is that you're trying to do.

So, I literally mean just writing that out in like plain text. Um and then this agent is either going to say, "Yes, everything is done. everything's satisfied. It's safe to close the issue. Uh in which case it can close close the issue from the terminal. So I never need to leave leave this view or it's going to say no this hasn't been completed. This hasn't been done properly. Duh. And you would just copy the response, paste it back to the dev agent to go and and complete all that stuff.

When the dev agent says it's done, come back to the QA agent, say the dev agent now said it's done. Can you please go and check again? Dev agent will check. And that's the kind of loop that I do. So each of these um is why I have them on the same line. they're working in parallel um not not not at the same moment but if this is working on GitHub issue 55 this will also work on 55 and so on. So there's a kind of pipeline that that they they go through and the dev agent also doesn't have permission to close GitHub issues.

Only the QA agent or the project manager has the uh ability to to um to be be able to do that. And this I have found to be like I've been like fine-tuning how to um how to do like vibe coding um I I hate the word vibe coding but let's just just go with it. How to like do vibe coding um for a long time and this is the best workflow that I have have found by a long long way. The other reason it's so good is because you're not needing to retain all of like the tasks in in your mind.

Uh everything is being put on on GitHub issues. The reason I also like GitHub issues instead of, you know, you could technically do like just issues locally, just like making files there. Um, but it's because you can create issues from from your phone as well. So, if I'm like out and about, um, you can actually just remotely come up with with an idea. You can also tag Claude in the issues as well, um, remotely and have it work on stuff for you.

Um, and also it has a direct integration with with Linear. So, Linear is like a a software project management tool, but it has a direct integration to GitHub. So, it it kind of has a two-way sync. And I'm a big fan of uh of linear. Um, and the other great thing as well is if you're working on multiple different projects, you can just have like a new window. Uh, so like this is like my like Reddit algorithm, but I can have another one for like another side project that I'm working on, and you can just easily like flick through these um just from one one to another and have three agents working on each project at at the same time.

Um, and yeah, I just I find it to be the the best way of working with these with like the least amount of headaches. If you don't have these QA agents, I find that you can get stuck in like a loop arguing back and forward with uh with the dev agents saying like, "No, that's not quite what I wanted. It's not completed to my satisfaction, but they've got it in their context that that it has been completed." And you just end up wasting a lot of time.

So, I find this to be like a really um a really good way to work. You can actually have it as part of the same workflow of you can tell Claude to launch a sub agent of a QA testing agent every time. And I have tested uh tested doing that as well. The only downside of that is it's always going to be a Claude agent. So you can't do what I mentioned of having Claude uh Opus 4.5 doing the actual coding and then have like GPT or Gemini doing the actual code review.

One final thing to mention is on the uh project management agent, it makes sense to use a model in my opinion that has a very high context window. Now we've actually seen big improvements in the context window of all of the models, but previously I used to only use Gemini 2.5 uh before 3 was released because it had a 1 million context window. So it was able to kind of understand your whole codebase and get a really good understanding of everything going on w within the project.

Um, and it could incorporate that into when it it writes out the tasks for the dev agents, it already has in its context everything that goes on within the um within the project. Whereas if you were doing it on like a agent with a much smaller context window, the the just that it's just not able to retain that much information within its memory. Um, yeah, I think that's that's everything. I hope that this was like uh something interesting and and insightful and a little bit different.

Uh if you do know of other people that are building out similar things to this. Um I'm not interested in like you know I've seen some tools out there where they're trying to sell sell a service and they're doing like 20 or 40% returns. It it's not interesting. Like I think it needs to be like 60 to 80% returns. Uh where it starts to to get interesting. So yeah, if you're working on something similar to this, uh I would be interested to uh to exchange some uh