What Do Mycroft Fellows Do? | Building AI-Powered Investment Intelligence | Professor Bear Explains
Professor Bear explains how Mycroft Fellows start by building small n8n financial components, then can grow into the four-layer agentic architecture behind an open-source investment intelligence system.
Most explanations of agentic AI in finance start with the architecture diagram. Professor Nik Bear Brown starts somewhere more useful: with the small, concrete thing a new Fellow actually builds in their first weeks on the Mycroft project, and works outward from there toward why that small piece matters.
Starting small: financial components
Mycroft is an agentic AI project, and like the other Humanitarians AI projects, new Fellows begin by building small components rather than jumping straight into the full system. For Mycroft, those components are specifically financial. A Fellow might build something that tracks congressional trades, since members of Congress have to register their trades, and it becomes genuinely interesting when five of them are all buying or selling the same thing at the same time. The same logic applies to insider trading activity. Other components look at market sentiment, unusual trading activity, or simply pull SEC filings and structure that data into something usable. These are the basic building blocks: small, single-purpose tools that gather one specific kind of financial signal.
Built on n8n, extended for orchestration
These components are built using n8n, an open-source automation tool. The open-source part matters specifically because it means Fellows can extend the tool and access its code for educational purposes. Mycroft extends n8n somewhat further with what Professor Bear calls the orchestration layer, the coordination system that lets individual components work together as actual financial agents rather than staying isolated, single-purpose scripts.
Why agentic AI is different from classic reinforcement learning
To explain what makes the orchestration layer notable, Professor Bear draws a comparison to old-school reinforcement learning, which has been around for a long time and remains genuinely powerful. It's what drives a Tesla's self-driving behavior and what powered chess-playing systems like Deep Blue. In a classic reinforcement learning setup, the environment is whatever the system perceives, camera input for a car, pixels for a game like Pac-Man or Pong, often captured across two or three frames to give the system a sense of time. Actions are whatever the system can do in response, moving a paddle, turning a wheel, hitting the brake. The state is a specific instance of that environment at a given moment, and the agent makes decisions based on it.
Agentic AI shifts this picture. In agentic AI, people often use language models, frequently without other supporting math, to make decisions directly. Professor Bear is clear that the language model itself is very powerful and effectively drives the agentic AI revolution, but it shouldn't be treated as the only piece of information worth using. One thing agentic systems share with traditional agents, though it's especially common here, is the use of triggers: something happens, and that event causes the agent to act.
A live example: unusual activity in Tesla and Meta
Professor Bear grounds this with a real, recent example rather than a hypothetical. In one week, Mycroft's agents detected something unusual with Tesla: in overnight trading, the stock dropped from around 333 to 300. Based on their models, the team judged this as a significant undervaluation and responded by buying a batch of options, an unusual move since options carry real risk, alongside a straightforward limit buy at 300. Because the move was judged highly unusual, the price began recovering quickly, climbing back toward 330, and the team placed limit sells at 330 and 327. Over the following week, the price kept oscillating, and the team sold in the high 320s and bought back in the high teens repeatedly as the pattern continued.
A similar signal fired for Meta. The stock had been around 715, then dropped under 700, prompting Mycroft's agents to recommend buying. Later the same day, right around the 4pm close, Meta gained roughly 12% after hours, a jump large enough that Professor Bear notes he would have expected to see it all over social feeds, and didn't, which itself is part of what made the signal valuable. The agents' recommendation to buy Meta around 698 to 699 meant the position gained roughly 80 points per share after hours on the same day.
Why this matters: humans don't have time to watch everything
The point of walking through these examples isn't the trades themselves. It's what they demonstrate about the value of agentic systems in a genuinely event-driven, time-sensitive domain like finance. A human being doesn't have time to notice that five senators all decided to sell the same stock on the same day, or to catch a 12% after-hours jump the moment it happens. What these tools do is surface the signal: something unusual happened, and that's the point at which a human needs to look. You have a life, you want to work out, relax, and do other things, but when something like this happens, you want to know about it. That gap, between constant market-wide attention and an actual human schedule, is exactly what agentic AI is built to close.
The path from Fellow to advanced contributor
The initial ask for a new Fellow is modest: build one small tool related to finance, whether that's gathering insider information, gathering stock information, gathering sentiment, or detecting unusual patterns. If a Fellow does good work, there are two directions to grow from there. One path continues building additional tools or works on the orchestration layer itself, extending traditional quantitative finance approaches, portfolio theory, risk management algorithms, financial forecasting models, into agentic workflows that adapt to changing market conditions. A second, more advanced path leads toward the deeply mathematical side of the work, represented by the Popper project, where questions get genuinely quantitative: how do you fundamentally and statistically pick a strike price, and how unusual does something have to be before it's unusual enough to act on.
This project is explicitly framed as a good fit for anyone with a real interest in finance who also wants exposure to agentic AI, whether that means working toward a job at a hedge fund, an investment bank, or simply building better tools for managing personal investments. Because the underlying tools are open source, a Fellow who builds here can bring the code into their own hedge fund, ETF, employer, or personal investing setup.
Key takeaways
- Mycroft Fellows begin by building small, single-purpose financial components: congressional trade tracking, insider activity, sentiment analysis, unusual activity detection, and SEC filing data.
- Components are built on n8n, an open-source automation tool, with an orchestration layer that extends it to coordinate components into working financial agents.
- Agentic AI differs from classic reinforcement learning by leaning on language models for decisions, though Professor Bear stresses that language models shouldn't be the only information source used.
- A real example showed Mycroft's agents flagging unusual Tesla and Meta price moves and acting on them with limit buys, limit sells, and options, illustrating how the system surfaces signals humans wouldn't have time to catch themselves.
- Fellows who do good work can progress toward extending the orchestration layer or toward the more mathematically advanced Popper project, covering questions like strike price selection.
- All of the tools built are open source, so Fellows can reuse the code in their own investing, in a future employer's systems, or in a personal project.
Who this is for
This is for anyone considering the Mycroft Fellows program who has a genuine interest in both finance and agentic AI, whether the goal is a career in investment finance or simply building better personal tools for tracking unusual market activity. Humanitarians AI is a nonprofit, and this video is part of a series meant to help prospective Fellows choose the project that actually matches their interests.
Full transcript(auto-generated, with timestamps)
[0:02]Hey, it's Professor Bear here again. I'm gonna talk about the projects we have at Humanitarian AI. This is Microoft. Like the other projects, this is an angentic AI project. So if you're going to be a Micro Fellow fellow feller, uh you're going to uh start by building little components. So components for um Microoft are financial. You might write a component which looks congressional trades because congressmen have to register their trade if maybe five congressmen at the same time are all buying or selling the same thing. That's interesting. Same thing for insiders. Uh we have um things that look at sentiment. We have things that look at unusual activity.
[0:48]We have things that just basically get you know SEC filings and that data. So those are the basic components. We use a tool called NAN. Uh the reason why you say n it's a great tool. It's sort of open source that mean that we can extend it and we get access to the code uh for educational purposes. Um, and it's a wonderful framework and we're going to extend a little bit and and this is the orchestration layer and I'll talk about that in a second. But if you start off as a fellow, you're going to be building a little component, little financial component that we can
[1:24]Use with the other components and make little financial agents. So what about this orchestration there? So old school ed reinforcement learning, it's been around for a long time. And it's very powerful. It's what makes a Tesla self-driving car drive. It makes, you know, things like all of these these chess playing um algorithms, deep blue, all that stuff work. So, how does this work? So, for example, with a car, your environment is your cameras. If you were playing a Pac-Man game or a Pong game, your environment would be the pixels on the screen. Typically, two or three shots so we can get a time aspect of it.
[2:05]If you're playing a pong game, your actions are moving the paddle up and down. If it's a car, your actions are turning the steering wheel or hitting the brake, that kind of thing. Your state is a particular instance. For example, what is in the camera at that moment of a car. That might be a red light, that might be a green light, that might be people coming across the street, whatever. If it's a game, it would be the particular the pong ball was coming right at your pedal or whatever it is. And the agent makes decisions. What's happened in angentic AI is a little different.
[2:42]So in intentic AI people often use the language models and so often the language models without other math to make the decisions. They just say hey you know do this do that with a language model. Our approach is the language model is very powerful. They they essentially drive the angentic AI revolution, but they're not the only bit of information you should use. Other things that are different with with uh angenic AI rather than traditional agents are triggers. Although triggers are used in regular AI as well, they're very common. So something happens and you know you start your car it cause causes the agent to start a a big event
[3:24]Happens in the stock market. Uh something happens. So let me talk a little bit about what those events are. So this past week uh our agents detected something very unusual Tesla. So and overnight trading Tesla went from like 333 around there 332 to 300 overnight. Because of that, we thought it was seriously undervalued based on our models. And we bought a bunch of options, which is very unusual for us because options are risky. And I'll do other uh videos on options and when to use options, when not to use options, etc. We also just used a limit buy and bought a bunch of Tesla at 300. Very quickly,
[4:07]Because we believe this was highly unusual, it started going back up. It actually went up to 330. We had limit sells at 330 and 327 and over the past week it's just been going up up and down. Tesla is an amazing stock. It'll literally go we will sell it at 330, 325, 327. The next day we'll buy it at, you know, 318, 319. Then the very next day we'll settle sell it again at, you know, 325, 326, 327. If you get deeper in this, we'll get into the math of how we pick our numbers. But if you are going to to start um this stuff, you're going to build an
[4:47]Agent for example just to look at unusual activity. Something is weird is happening with Tesla. Our agents also told us to today to buy Meta even though Meta literally like a day or two ago was 715. Then it dropped under 700. So we we got a bunch of stuff. And then something amazing happened today at the end of the literally at the end of you know trading four four o'clock it gained 12% after hours. So we this is this price is actually all this is what we we got our Tesla around I mean our meta at you know I think 698 699 it went up 80 today. This would be fun.
[5:29]I don't know what's happening with meta. Something happened with meta. That's a 12% jump after hours. And this is where the angantic AI is so powerful. There are a lot of stocks. And if something happens like this 12% jump after hours, I I don't know what's going on with Meta. I haven't looked at the news. I actually had I would expect a jump this big would show up all over my, you know, LinkedIn feed and other feeds. It hasn't. Something happened, though. And we can tell that something happened because it went up 80 points after hours in a very short time like an hour after
[6:07]Hours. So I'm guessing they did some announcement and sometimes these announcements are timed after hours because of this something happened. And so these these are why these agents are so powerful is you don't have time as a human being to know that four senators decided to you know or five senators all decided to sell this stock. Uh you have a life. Um same thing with all these other things. But what what the tools allow you to do is this is unusual. When it's unusual a human being needs to look at it. Something is happening here. Something is still happening with Tesla. And Tesla's a very
[6:47]Unusual one because of Musk. I think he just you know people are not rational when they they they come to to Tesla which you can take advantage of as a as an investor. So back to the micro what you will do initially is just build a a tool something useful gather inside information gather stock information gather sentiment detect unusual patterns you know Tesla dropped 10% after hours that goes on your watch list you need an email saying you know get up before the market opens and start looking at Tesla to start thinking about how you want to take advantage of of this. Same thing right now with meta. This is real time.
[7:36]I don't know what's going to happen with Meta tomorrow, but that's unusual. 12% after hours. And fortunately, our bots told us to buy some meta, which we did. Wish we had bought more um because of 80 $80 per share today after hours. And so this is the power of of these engineic systems in finance. Finance is very event driven, very timesensitive. Uh you you know the the weak old Tesla thing is you know you can still sort of get the trail end of it but you needed to be there at the time. You need to be there at the stock market tomorrow dealing with Meta
[8:22]Because something happened today with Meta which was highly unusual. it whacked from 715 down to 700 and then back up over 780. So, so this is what these things do. You have a life. You don't uh you know, you don't want to spend your day just looking at ticker symbols. You want to you know work out and you know relax and do things but when those things happen you want to know about it. And this is the power of engineic AI. So as a fellow if you're interested in this world of finance and trading, I happen to have an MBA and a masters in finance and we have
[8:58]Other experts who work with us uh on these things. Uh but what you are initially going to do is build a little tool related to finance. And so if you're if you're really thinking of working for a hedge fund or working for Goldman Sachs, working for a bank, this would be a good good thing for you. It's less sort of creative than than the other projects like the Microsoft. you can make videos and and little commercials and things like that. Popper is deeply mathematical. So, if you're interested in fundamental fundamental, the math in finance is pretty simple. Um, but the the the popular math is
[9:35]Really interesting. And again, you'll have some math here because how do we pick the strike price? You know, how do we do that fundamentally and statistically? But that's more advanced. So the the way it works is you join the project you you build a little tool. If you do a good job then you can either continue building little tools or you can get promoted to either dealing with a math you know how do we decide a limit price? How do we decide to place an option which is a highly risky thing which we rarely do unless it's highly unusual. How unusual has to be unusual?
[10:11]Those are all mathematical questions and then start thinking about improving the angentic architecture of basic reinforcement learning for finance. So I hope this gives you a sense of what it's about. Like most of the projects, it's centered around angentic AI. And with any of the projects, you will learn angentic AI. But if you have a more of a passion for finance, you want to get a job in finance, it helps to or just, you know, want to make money in the stock market, um, this would be a good project for you. So I I hope this gives you a sense of what the micro project is. Uh why you
[10:52]Might want to be involved with this one rather than another angentic pro project. But this is you know this is finance. It's a finance thing. So uh I'll do my usual thing. Uh humanitarians AI is a nonprofit. Uh most people do not donate to nonprofits and that's okay. But if you like this, share it, you know, ring the bell, do that things. It helps us uh and I think these are useful useful tools that we build uh open source tools. Anybody can use this is open source. So if you work on this project, you can start your own little hedge fund or ETF or whatever you want the code.
[11:31]You can bring it into the company that you work with. You can, you know, just use this for your own personal finance of if you have some stock market investing and you want to be alerted when highly unusual things happen or when seven congressmen all buy buy the same stock at the same time, then this would be a nice nice project to get involved with. So hopefully that gives a sense of what this is about. I will be making videos on all of our projects. And the purpose of these videos is so you can make a rational choice of if you want to get involved, what project that
[12:05]You want to get involved in. And we we our basic philosophy is we want you to be passionate about the project. Our experience is if you do not care, you do not do a good job. So, we don't force you into a particular uh project. But if you do a project, we expect you to commit to that project. We have simpler projects than these and I'll make videos on those. If you just want OPT and want to do something very simple, we have those projects as well. But this is project for somebody who has a real interest in finance and a real interest in AI. Okay, that's it. I've
[12:40]Already gone way too long. Uh but it is what it is. Take care.
More from Mycroft Financial AI
2:59What You'll Do If You Join Mycroft | Financial AI OPT-Eligible Open Source Project
10:00Building AI News Monitoring Agent with n8n & FastAPI | Mycroft | Nerd Stuff with Humanitarians AI
3:48What is Agentic AI in Finance? Meet Mycroft
5:36Project Mycroft: Building a Structural Enforcement Layer to Stop Silent AI Failures
3:43Mycroft AI
2:44