Show HN: I built an AI agent that helps me invest
github.comA while back, I built a simple app to track stocks. It pulled market data and generated daily reports based on my risk tolerance. Basically a personal investment assistant. It worked well enough that I kept going.
Now, the same framework helps me with real estate: comparing neighborhoods, checking flood risk, weather patterns, school zones, old vs. new builds, etc. It’s a messy, multi-variable decision—which turns out to be a great use case for AI agents.
Instead of ChatGPT or Grok 4, I use mcp-agent, which lets me build a persistent, multi-agent system that pulls live data, remembers my preferences, and improves over time.
Key pieces: • Orchestrator: picks the right agent or tool for the job • EvaluatorOptimizer: rates and refines the results until they’re high quality • Elicitation: adds a human-in-the-loop when needed • MCP server: exposes everything via API so I can use it in Streamlit, CLI, or anywhere • Memory: stores preferences and outcomes for personalization
It’s modular, model-agnostic (works with GPT-4 or local models via Ollama), and shareable.
Let me know what you all think! Well the most important information is missing: what was your ROI ? :D I'm going to assume this project is at best a few years old, so return over index is more likely to be an indicator of overbetting (taking on too much risk) than a performance indicator. Backtesting would be more useful. Of course, LLMs cannot be backtested since they know the past. This system is impossible to test. I would be hesitant to trust it. >> I'm going to assume this project is at best a few years old, so return over index is more likely to be an indicator of overbetting (taking on too much risk) than a performance indicator. Best metric here would be a Sharpe Ratio and drawdown details.
https://en.wikipedia.org/wiki/Sharpe_ratio > This system is impossible to test. I would be hesitant to trust it. Disagree. Best test would be a paper test going forward, audited by a common platform. Any forward test depends on market assumptions. I'm not a practitioner, and as a layperson I wouldn't know whose assumptions to trust. it looks like the code was committed a week ago. I do not see an iterative history on the code I didn’t build this to beat the market...it started as a way to reduce my own decision fatigue and make sure I wasn’t missing obvious signals. That said, it’s more of a research agent than a trading bot. For ROI, it’s not auto-trading, but it has helped me avoid a few bad calls and made my reports more consistent and thorough. Sounds interesting -- can you share some real examples? In context of investing, what does "It worked well enough" translate to? Yeah I’d like to know this too. I have also done well (enough) with investing by literally sitting on my index fund investments. I'd be curious also. Is it possible to try your real estate AI agent? I've built a real estate data API (stream.estate) and would be very much interested to see how it works; maybe we can add some features to make that kind of project better. Hate to be that guy but I'm not sure how this is any better or more reliable than something like FinViz or NeighborhoodScout. Great question and those are solid tools! Where this differs is the orchestration. Instead of switching tabs or manually checking five sources, this bundles everything into one interactive, LLM-assisted report, plus it remembers your preferences, investment style, and context. You can even swap out GPT for a local model if you’re privacy-conscious or budget-sensitive. Because it has AI. The latest, shiniest shiny object! The ideal design for such an agent is that it would tell you to put all your money in a Betterment account, and then if you try to do anything other than that it should give you electric shocks until you stop. (Vanguard if you don't benefit from tax loss harvesting.) Is that going to beat FXAIX in a Fidelity account? The expense ratio is 0.015% and it tracks the S&P 500. The main benefit of something common and well supported like VIX is that it's easier to transfer between companies without selling it, which for a brokerage account is a taxable event. Fidelity has no expense funds or low expense funds, but if they change their mind or the company becomes bad for you, you can't move it to another provider without being taxed. If you will never leave Fidelity or if it's not taxable anyways (retirement) then it's fine, but if you think you might have to move assets then it's a potential risk unless other providers also support them. Being trapped is not worth the slightly lower expenses IMO. Thanks, that helps. In this case, it's my IRA money anyway, but that's good to keep in mind. If you're fine with 100% risk then probably not, but you want some low-risk assets depending on your age. And if you have high marginal tax rates like anyone living in CA does then a little tax optimization is worth it. Oh, also mutual funds like FXAIX are less tax-efficient than ETFs. Happy to elaborate on how ARC OS works —
It parses subjective input into logic trees with assumptions, conflict checks, bias flags, and reasoning trails. It’s symbolic only (no LLMs), designed for alignment auditing, law/policy frameworks, and decision explainability. If anyone wants an example, I can post a breakdown here. I think others are asking more about the real-world investment value of this, not the technical implementation. What actual trades were made by the user/creator? What was the ROI? How did profitability compare to their returns before using this tool? With today's LLM's it's easy for anyone to generate a 20-page "report" with a analysis about investments. But a report that, when followed, actually gives you above-average returns? No one has shown evidence of that yet. Sorry, I think I replied in the wrong thread.
My comment was meant for a different topic — feel free to ignore!