Somewhere along the way, I decided I needed a server.
Not a particularly exotic server. Just a small computer that could sit quietly in my house, running a handful of applications I’d been building in my spare time.
Naturally, I named it AndrewServer3000.
What began as a fairly practical exercise in organizing my digital life has evolved into something considerably more ambitious. I’ve been building a personal digital archive, experimenting with artificial intelligence, creating autonomous research tools, and developing a system for testing business ideas.
At some point, I realized I wasn’t just building individual applications anymore. I was building an ecosystem.
I’m not entirely sure when that happened.
First, There Were the Photos
I’ve been a photographer for decades. Consequently, I have decades of photographs scattered across hard drives, optical discs, and assorted storage devices. Some are carefully organized. Others are effectively archaeological discoveries.
I wanted a way to search and browse that collection without having to track down the original files every time.
So I started building my own digital asset management system using Python, Flask, and SQLite.
The application catalogs photographs, generates smaller viewing copies, extracts metadata, identifies potential duplicates, and uses facial recognition to help organize people across the collection. I’ve also been adding video support and building an Android companion application for uploading media directly from my phone.
It has a ridiculous name that I’m going to spare you here. Suffice it to say that the entire family of applications follows a naming convention that probably wouldn’t survive a corporate branding review.
The important part is that it works.
And the more I worked on it, the more comfortable I became building applications outside my usual world of website development.
Then I Started Thinking About Memory
One of the things I find simultaneously fascinating and frustrating about AI is its relationship with memory.
A conversational AI can be remarkably capable, but that doesn’t necessarily mean it can reliably recall something you discussed months ago, much less retrieve an exact passage from a document.
I’ve accumulated years of conversations with ChatGPT, including discussions about software development, writing, research, and more than a few spectacularly unsuccessful experiments.
I didn’t want all of that history to disappear into a giant pile of exported files.
Enter MemPalace, a system I’ve been experimenting with to preserve and retrieve information from those conversations and documents.
The goal is more ambitious than simply searching old chats. I want to be able to ask questions about previous projects, retrieve original source material, and distinguish between what an AI actually knows from a document and what it merely remembers discussing.
That distinction matters.
For example, I’ve been experimenting with having AI analyze a novel manuscript. I want its assessment of the manuscript itself, not an opinion contaminated by months of conversations in which I might have suggested what it ought to think.
Working with MemPalace has also exposed some interesting limitations in how imported conversations are represented and retrieved. It’s been a reminder that storing information and making it reliably accessible are two very different problems.
Meet Anton
Around the same time, I started wondering whether AI could do more than answer questions and help me write code.
Could I build a system that takes an idea, researches its feasibility, proposes an experiment, and helps me determine whether it’s worth pursuing?
That became Andrew Ventures, my personal experiment in building a tiny, AI-assisted venture studio.
At the center is an application called Anton.
Anton is built with Python, Flask, and SQLite. It provides a structured workflow for capturing ideas, gathering research, developing hypotheses, and designing experiments.
The idea is to stop treating every interesting thought as either a brilliant business opportunity or another project destined to collect dust.
Instead, each idea gets evaluated.
Is there a market? What evidence supports it? What’s the cheapest meaningful experiment? How much time and money would it require? What would constitute success?
I’m also exploring how autonomous AI agents might carry out bounded research and development tasks while Anton keeps track of the results, expenses, and decisions.
I’m particularly interested in the distinction between an AI system that recommends doing something and one that can actually execute a well-defined task, report what happened, and be held accountable for its results.
The goal isn’t to hand over my credit card and let a robot start companies while I sleep.
Although, admittedly, that would be entertaining.
And Then the Robots Started Arguing About Stocks
One of the first experiments I began exploring for Andrew Ventures involves TradingAgents, an open-source framework that uses multiple AI agents to analyze financial markets.
The concept is fascinating.
Rather than asking one AI model whether a stock is a good investment, the system assigns different analytical roles to different agents. Researchers examine market conditions and technical indicators. Bullish and bearish analysts debate the evidence. Other agents assess risk and help formulate a decision.
I’ve been experimenting with running this framework using Qwen models and evaluating the reasoning it produces.
There’s something wonderfully absurd about watching artificial intelligence argue with itself over Microsoft stock.
Of course, producing an impressive-sounding investment thesis isn’t the same thing as producing a profitable investment strategy.
That’s why I’m treating this as a research experiment, not an invitation to turn my retirement savings over to a committee of chatbots.
The interesting question is whether these systems can generate useful, measurable insights when evaluated against ordinary investment benchmarks.
And if the answer is no, that’s useful information too.
Bringing It All Together
Which brings me back to AndrewServer3000.
I recently acquired a MINISFORUM MS-01 mini PC to serve as a dedicated home for these experiments.
The long-term plan is to bring several pieces together:
- My digital archive: A searchable collection of photographs, videos, and their metadata.
- MemPalace: Long-term retrieval of conversations and source documents.
- Andrew Brain: A personal knowledge environment built around AFFiNE and related tools.
- Anton and Andrew Ventures: A central place to manage ideas, research, experiments, and outcomes.
- TradingAgents: A sandbox for exploring multi-agent financial analysis.
- AI orchestration: An experimental layer that can coordinate bounded tasks across different tools.
Not everything is finished, and some components are still being developed or evaluated. That’s part of the fun.
The larger vision is to have a collection of applications that can share useful information and support one another, rather than a dozen unrelated experiments running on my laptop.
Why Am I Doing All This?
That’s a fair question.
I’ve spent much of my professional career developing and maintaining websites and web applications. I enjoy solving problems, particularly the ones that require me to learn something unfamiliar.
But there’s something different about working on personal projects.
Nobody has written a requirements document. There isn’t a committee reviewing the color of a button. I don’t have to justify every experiment with a projected return on investment.
If I want to spend an evening figuring out how facial recognition works, I can.
If I want to see whether five AI agents can have a productive argument about a stock, I can do that too.
And if I want to name a computer AndrewServer3000, there is absolutely nobody around to stop me.
I’ve also learned an enormous amount in a relatively short period. Python, Flask, SQLite, Android development, local AI models, agent orchestration, and the practical realities of getting all these technologies to cooperate.
AI has made some of that exploration more accessible, but it hasn’t eliminated the need to understand what I’m building. There are still architectural decisions to make, bugs to investigate, security considerations to address, and plenty of occasions when the generated code doesn’t do what it’s supposed to.
The most valuable part isn’t that AI can help write code. It’s that I can explore ideas that would previously have required far more time, specialized knowledge, or outside assistance.
I’m not trying to become the next Silicon Valley startup founder. If one of these experiments eventually generates some income, terrific.
But even if none of them ever makes a dime, I’ve already gotten something valuable out of the experience.
I’m learning. I’m building things that interest me. And I’m having a ridiculous amount of fun doing it.
Besides, I’ve got a server to justify now.