Latest / Deep Values: Value Investing Deep Dives / Build Your Own Hedge Fund with AI Multi-Agents - DeepValueIntelligence (OpenSourced)
Transcript
- 0:00You know, I was looking at My Portfolio the other day, just
- 0:03staring at this one ticker that was deep in the red, and I had
- 0:06this realization. Investing is a really, really
- 0:09lonely sport. It is.
- 0:11It's just you against the market and the market has, well, a lot
- 0:14more resources than you do. Right.
- 0:16And it's not just lonely, it feels like it's getting almost
- 0:19impossible for a normal person to actually win.
- 0:22I mean, to do this properly, to really beat the market, I feel
- 0:25like I need to be a forensic accountant, a technical analyst,
- 0:30a risk manager, to tell me I'm being an idiot and a news junkie
- 0:35all at once. And I'm supposed to do all that
- 0:37at 10:00 PM after my actual job? That is the classic dilemma.
- 0:41You're trying to replicate the work of an entire hedge fund,
- 0:44you know, with 50 analysts and expensive terminals with just
- 0:46one brain and what, a few hours a week?
- 0:50Exactly, and you end up wearing too many hats and one of them
- 0:53slides down over your eyes. Yeah, maybe you're great at the
- 0:56math, but you completely miss the market sentiment.
- 0:58Totally. And when we try to use AI for
- 1:00this, we just, you know, open a chat bot and ask, hey, is this
- 1:04stock a buy in that spits out this this paragraph that sounds
- 1:07super confident but says absolutely.
- 1:09Nothing. It's the fortune cookie problem.
- 1:11It sounds profound, but it applies to everything and
- 1:13nothing, right? But today we're doing a deep
- 1:17dive on something that I think completely changes that whole
- 1:20dynamic. It's an open source framework
- 1:22called Deep Value Intelligence, and what grabbed me is that it
- 1:27doesn't try to give you an answer, it builds you a staff.
- 1:30And that's the key distinction right there.
- 1:32This isn't a tool. You just ask a question.
- 1:34It's a framework that builds you a virtual investment firm.
- 1:37AI agents take on these specific jobs.
- 1:39You get a boss, employees, even a department that just exists to
- 1:43say no. A no department.
- 1:44I definitely need one of those. My yes department is way too
- 1:47active. So D value intelligence or DVI
- 1:49is the topic. But you mentioned there's a step
- 1:510 right before we unleash the AI army.
- 1:54A crucial first step. You have to read the room.
- 1:57You can have the best stock tick in the world, great earnings,
- 2:00great product, but if the macro environment is just falling
- 2:06apart. Your stock is going down with
- 2:07the ship. It's going down with the ship.
- 2:09You have to know the weather before you set sail.
- 2:11So how do we get that weather report?
- 2:14Because just looking at the S&P 500 doesn't always tell you the
- 2:16whole story. To track total market
- 2:18temperature, welcome to search and download the iOS app Buffett
- 2:22Indicator from the Apple App Store.
- 2:24It gives you that really high level view, specifically the
- 2:27ratio of total market cap to GDP.
- 2:29So is the market hot or cold compared to the actual economy?
- 2:32Exactly. It sets the stage for everything
- 2:35else. OK, so check the temperature
- 2:37with the app first. Don't buy a house in a
- 2:39neighborhood that's on fire. Got it.
- 2:42Then we deloy the squad. Our mission today is to unpack
- 2:46how this DVI framework really works.
- 2:49We're going to look at how it replaces that one simple prompt
- 2:51with a whole team of agents that, well, they actually debate
- 2:55each other. They get pretty heated too.
- 2:56And they produce a final stance. Advocate, watch or avoid.
- 3:00And just to be clear for everyone listening, this is open
- 3:02source. It's a Python framework you can
- 3:04run on your own machine. It's all transparent, you can
- 3:07see the gears turning. It's not some black box website.
- 3:10So let's start with the basic problem.
- 3:13Why is that one shot prompt so bad?
- 3:16It's so convenient to just ask a chatbot.
- 3:18Why is that so dangerous? It's dangerous because large
- 3:21language models, well, they're designed to be agreeable.
- 3:24When you ask should I buy Apple? It's just doing probabilistic
- 3:28word association. It might pull APE ratio from two
- 3:31years ago from its training data.
- 3:34Mix it with a headline from yesterday.
- 3:35I completely ignore the chart. Exactly because it can't see it.
- 3:39But the biggest issue is it lacks state.
- 3:41State, OK, break that down for us.
- 3:44It doesn't have a working memory of its own thought process.
- 3:47It doesn't stop and say, wait, I found this great revenue number,
- 3:49but I need to check that against the huge debt levels I saw a
- 3:52second ago. Deep Value Intelligence forces a
- 3:55strict pipeline. It mimics a real research firm
- 3:58through an agentic workflow. I've heard that term agentic
- 4:01being thrown around a lot sounds very spy movie.
- 4:04It's simpler than that, but yeah it's cool.
- 4:06Instead of 1 AI trying to do everything, you break the task
- 4:10into steps and you assign specialized AI agents to each
- 4:14step. First gather evidence, then
- 4:17argue about the evidence, then a reality check, then a final
- 4:22decision. And it creates a paper trail or
- 4:24I guess a digital trail. Traceable reasoning.
- 4:28That is the huge take away here. The value isn't just the final
- 4:31answer, it's the audit trail. You can see why the value
- 4:34analyst loved it and why the risk manager hated it.
- 4:36You can literally watch the argument play out.
- 4:39OK. I want to meet the team.
- 4:40If I'm the CEO of this virtual firm, who's on my payroll?
- 4:44All right, let's do the roll call.
- 4:45First up, you've got the fundamentals analyst.
- 4:47The person in the back office with the thick glasses and
- 4:50stacks of reports. That's the one.
- 4:51This agent is all about the hard numbers.
- 4:54It does a full 10 year deep dive.
- 4:56Financial strength, profitability trends, revenue
- 4:58growth and it runs DCF. Scenarios discounted cash flow.
- 5:02The thing I always promise myself I'll do, but never
- 5:04actually open Excel for. Exactly because it's tedious.
- 5:08This agent does it instantly. It's not looking at the stock
- 5:11price, it's looking at the business.
- 5:12What is this cash machine actually worth?
- 5:15OK. So that's the foundation, but
- 5:16then you have a more specific persona, the value analyst, and
- 5:20the code literally calls this one the Buffett bought.
- 5:24It is. This agent is programmed to
- 5:26think like Warren Buffett. It doesn't care about hype or
- 5:30buzzwords. It's looking for moats, you
- 5:32know, durable competitive advantages.
- 5:34It looks at management, quality and most importantly, that
- 5:38margin of safety. Is this dollar bill selling for
- 5:41$0.50? Precisely, if a stock is trading
- 5:44at 50 times earnings, the Buffett bought is going to have
- 5:47a problem with it, even if the whole market loves it.
- 5:49But you can't just run a portfolio on deep value alone.
- 5:52You'd have missed almost every major tech trend.
- 5:54You'd have missed Amazon, Netflix.
- 5:56Which is exactly why you also hire the growth analyst.
- 5:59This one mimics the style of investors like Peter Lynch,
- 6:02Stanley Druckenmiller, Philip Fisher.
- 6:05The heavy hitters of growth, the rocket ship guys.
- 6:09Right. This agent doesn't really care
- 6:11if the stock is cheap today. It cares about the total
- 6:14adjustable market, the Tam. It's looking for inflection
- 6:17points for runway. Can this company triple in size?
- 6:20So you have the Buffett bot saying it's too expensive and
- 6:23the Lynch bot saying but look at the potential.
- 6:27You can already feel the conflict building.
- 6:29And that conflict is where the real insight comes from.
- 6:31If they both agree, that's a powerful signal.
- 6:34If they disagree, you know exactly what the core tension of
- 6:37the investment is. But we're still missing the
- 6:39charts. I've been burned so many times
- 6:41buying a great company while its stock is just falling off a
- 6:44Cliff. And for that you have the market
- 6:46analyst. This agent just looks at the
- 6:48technicals, make out, DRSI, Bollinger bands.
- 6:52It answers one simple question. Is the trend our friend?
- 6:56You might love the fundamentals, but if the stock's in free fall,
- 6:59this agent waves a big red flag. And finally, there's the one I
- 7:03find kind of hilarious but also necessary.
- 7:06The social and news analysts. The vibe.
- 7:08Check it's his. Literally a vibe.
- 7:10Check. It scans Reddit, social media,
- 7:12chatter, all of it. It sounds silly, but in the age
- 7:14of meme stocks, you have to know if the vibes are toxic or if
- 7:18there's some huge macro event like a war or a new regulation
- 7:22that you won't see on a balance sheet from three months ago.
- 7:25So we have this team. They've done their homework.
- 7:28In a normal world they just e-mail me a bunch of PDFs, but
- 7:31here they enter the war room. And this is the core of it all.
- 7:35In a normal chatbot interaction, the model just agrees with
- 7:38itself. It's a sycophant.
- 7:40Here, DVI introduces the bull versus bear researchers.
- 7:44My favorite part? You're forcing the AI to fight
- 7:47itself. It's called adversarial thesis
- 7:49testing. The bull researcher takes all
- 7:52that data we just gathered and builds the best possible case
- 7:55for buying the stock, and the bear researcher takes the exact
- 7:58same data and tries to tear that case to shreds.
- 8:01So the bull says revenue is up 20% and the bear comes back with
- 8:04yeah, but margins are down 5%. Exactly.
- 8:06They are locking horns and watching.
- 8:09This whole thing is the research manager.
- 8:11This agent has to listen to the debate and synthesize a plan.
- 8:15But, and this is critical, the manager doesn't make the final
- 8:19call. No, because even a manager can
- 8:21get swept up in a good story. You need the no department.
- 8:25You need the risk team. This is the final boss.
- 8:28Even if the research manager loves the idea, the report goes
- 8:32to the risk team. And the Risk team has its own
- 8:35internal debate. It does.
- 8:37You have a risky agent who might say go big or go home.
- 8:40You have a safe agent who's screaming about position sizing
- 8:44and then presiding over them is the risk judge.
- 8:47The risk judge sounds serious. It should be.
- 8:50The risk judge makes the final call.
- 8:52It produces the deep values commentary.
- 8:55It looks at the downside, the volatility and decides if the
- 8:58trade is even cleared for take off.
- 9:00It's the ultimate gatekeeper. I saw in the code there are
- 9:02safety controls because I can just imagine 2 AIS arguing
- 9:06forever. No, you're wrong.
- 9:06No, you're wrong. Right, an infinite loop of AI
- 9:09bickering. They have hard stops.
- 9:11There's a setting called Max debate arounds.
- 9:13It forces them to conclude you get 3 rounds to make your point
- 9:16and that's it. That is just so cool.
- 9:18It feels so much more, I don't know, human.
- 9:21It captures the messiness of a real decision.
- 9:23It captures the dialectic thesis, antithesis, synthesis.
- 9:27You're basically forcing the system to kill its own
- 9:30confirmation bias by making it argue the negative case.
- 9:34OK, let's pop the hood for the tech folks listening.
- 9:37I'm looking at the GitHub repo. This isn't just a simple script.
- 9:40How is it built? It's built on a library called
- 9:42Langgrah. OK, what is Langgrah?
- 9:45OA lot of people know Lang chain which is kind of a straight line
- 9:47step A then step B then step C. Langgrah is for building
- 9:51stateful multiagent systems. It's more like a flow chart with
- 9:55loops. It allows the system to say the
- 9:57risk team rejected this, send it back to the analysts for more
- 10:01data. So it's not a black box where
- 10:02you just throw data in and hope for the best.
- 10:04You can see exactly where the process is at any given moment.
- 10:07Precisely, it manages the state of the investigation.
- 10:10And there's another really smart feature I saw in the config
- 10:12file. They used two different speeds
- 10:14of thinking. Quick think and deep think.
- 10:17Right. Think of it like a junior
- 10:18analyst versus a senior partner. For the heavy lifting the tool
- 10:22based tasks like scraping data or finding APE ratio.
- 10:27They use a faster cheaper AI model.
- 10:29Something like Gemini 3 flash preview?
- 10:32But for the big judgement calls. For the synthesis for the risk
- 10:35judges final decision that's when they switch to the deep
- 10:38think model like Gemini 3 pro preview you spend the compute
- 10:42power where it actually matters on the cognitive heavy lifting
- 10:45it's very economical and what. About data, does this thing need
- 10:48to be online all the time? It's flexible.
- 10:51There's an online mode and an offline.
- 10:52Mode. Online mode is obvious, I guess
- 10:54it hits the live AP. Is Yep, Yahoo Finance, Fin hub,
- 10:57Google News, real time data. But the offline mode is actually
- 11:02really interesting for serious quants.
- 11:04And why would you want old data? Back testing.
- 11:06If you want to see how your staff would have performed
- 11:08during the 2008 crash, you can feed it cash.
- 11:11Historical data. You can test your agents against
- 11:13history. So you can see if they would
- 11:15have saved you from a disaster before you trust them with real
- 11:17money. You can essentially time travel
- 11:20your AI team and I saw it uses something called Chroma DB for
- 11:24memory. Yeah, what's that?
- 11:25That's for vector storage. It gives the agents a long term
- 11:29memory so it can retrieve context from past analysis
- 11:33instead of starting from scratch every single time.
- 11:35It's like they're filing cabinet.
- 11:37OK, so I run the script, the debate happens.
- 11:39What do I actually get at the end?
- 11:41You get a folder and inside that folder is a markdown report.
- 11:44I love Markdown, clean and simple.
- 11:46And inside you get sections in intelligence summary, all the
- 11:50individual analysts reports, the full debate history, you can
- 11:53read the whole argument and then the final stance.
- 11:56And the stance is one of three labels.
- 11:57Advocate, watch or avoid. Simple.
- 12:01But there's one last feature that really just blew my mind.
- 12:03This reflection module. This feels like pure sci-fi.
- 12:07It's the most futuristic part for sure.
- 12:09It's in a file called reflection dot pywire and it allows the
- 12:13system to do what they call continuous learning.
- 12:15How does that work? Software doesn't usually just
- 12:17will learn on its own. Well, this is different.
- 12:19After a decision is made and some time passes, the system can
- 12:23look at the actual outcome. Did the stock go up or down?
- 12:28It generates lessons from that outcome and stores them in that
- 12:31chroma DB memory. So.
- 12:33If the risk judge was too cautious and it missed a huge
- 12:36rally, it remembers that. Exactly.
- 12:38It stores that as a lesson. So the next time a similar
- 12:41pattern comes up, say another high growth, high volatility
- 12:44stock, it retrieves that memory. The system can effectively say,
- 12:48hey, last time we saw this, we were too scared and missed out.
- 12:51Let's adjust our thinking. That is wild.
- 12:54So it's not just running a script, it's actually evolving.
- 12:56It's developing a gut feeling based on experience.
- 13:00That's the idea, but I feel like I have to be the risk team for a
- 13:03second here. We need a reality check.
- 13:05OK, Yeah. Let's pour some cold water on
- 13:07this. What are the big limitations?
- 13:09First and foremost, the documentation screams this.
- 13:12It is not an execution engine. It doesn't trade for you.
- 13:15It's a research search tool. You still have to push the
- 13:17button which is. Probably a good thing, yeah.
- 13:19I don't want a Python script emptying my bank account.
- 13:22Definitely not second model non determinism.
- 13:26You can run the same analysis twice on the same day and get
- 13:29slightly different results. The AI's are creative, not
- 13:32calculators. So don't treat it like gospel.
- 13:35Treat it like a very, very smart second opinion.
- 13:38And finally, data lags. If the API's are down or the
- 13:42data is bad, the analysis will be bad.
- 13:44Garbage in, garbage out. The agents are only as smart as
- 13:47the information they can read. So it's not a magic money
- 13:50machine, it's a super powered research assistant.
- 13:52It's a productivity multiplier. It does the work of 6 analysts
- 13:56in 5 minutes and that frees you up to be the actual portfolio
- 13:59manager to do the real thinking. It really feels like we're
- 14:02moving past just using AI tools and into managing AI employees.
- 14:08Deep value intelligence is sort of democratizing the entire
- 14:11structure of a hedge. Fund.
- 14:12It brings it right back to what you said at the start, the
- 14:14loneliness of the individual investor.
- 14:16This gives you a team, but it raises a really provocative
- 14:19question for me. If the AI can analyze debate,
- 14:23manage risks, and even reflect on its own mistakes to get
- 14:26better, at what point does the human in the loop become the
- 14:30bottleneck instead of the guide? That's a heavy thought to end
- 14:35on. Are we the CEO or are we just
- 14:37the slow part of the circuit board?
- 14:39Something to think about. Well, on that slightly unnerving
- 14:42note, we do have to wrap up. But before you go and download
- 14:45this, I need to read the official disclaimer, so listen
- 14:47closely. Full deep value research reports
- 14:50can be found in the description links.
- 14:52All analysis is based on publicly available sources and
- 14:54may be subject to revisions or differing interpretations.
- 14:58The content is for educational purposes only and does not
- 15:01constitute investment advice. Conduct your own due diligence
- 15:04before making investment decisions.
- 15:06If you want to access more advanced contents, welcome to
- 15:09subscribe on Deep Values dot AI. We will share our latest
- 15:12researches and the most advanced AI agent tools for our
- 15:15contrarian friends. Don't forget to like this video.
- 15:17Much appreciated. Happy investing out there.
- 15:20We'll see you in the next deep dive.