I Built an AI Stock Research Tool Inspired by Wall Street

This AI stock research tool screens the top 100 stocks across the US and ASX markets, runs four specialist AI analyst agents across every one of them, and delivers a full research brief in minutes. On Wall Street, that takes an entire team. If you invest in stocks, even casually, you know the problem.

Properly researching a single company takes time most people don’t have. You need to check the chart, read the latest news, dig into the financials, and see what insiders are doing before deciding whether it’s even worth acting on. Most people skip steps. Most people trade on incomplete information.

Wall Street solves that problem by deploying entire teams. A large investment bank runs dedicated desks for exactly this: fundamental analysts who dig into company financials, technical traders who read price action, macro strategists who track news and economic events, and quant desks that monitor institutional money flows. They work in parallel, pool their signals, and a portfolio manager makes the final call so no angle gets missed before capital is committed.

I built Quant Trading OS to change that. It deploys four specialist AI analyst agents, one for fundamentals, one for technicals, one for news and macro events, and one for smart money flows, mirroring exactly the multi-desk structure the professionals use. It runs all four agents simultaneously across the top 100 stocks from the US market and the ASX 200, combines their signals using a market-aware scoring system, gates every idea through a hard-coded risk filter, and then asks Claude AI to write a full research brief: key price levels, what’s working for the stock, what’s working against it, and a suggested position size based on your capital.

The whole thing was built over 13 structured sessions using Claude Code, Anthropic’s AI coding tool, with written acceptance criteria for every session before a single line of code was written. This post covers what it does, how it works, and what the results look like.

The AI can recommend. The risk engine must approve. You make the final call. That’s the only design that makes sense.

Four AI Analysts, Running in Parallel — Just Like a Wall Street Desk

The core idea behind Quant Trading OS is simple: instead of one AI trying to do everything, the app runs four specialist AI analyst agents — each one focused on a different angle of the same stock, exactly the way a professional investment desk would divide the work. Here’s what each one does.

---

---

How the Signals Come Together — The Portfolio Manager Agent

Once all four agents have run, a fifth component, the Portfolio Manager Agent, combines their scores into a single composite signal. But it doesn’t treat every agent equally. The weighting changes based on market conditions, measured by the VIX (the market’s “fear index”).

Market condition	VIX level	Who gets upweighted	Why
Calm market	Below 15	Fundamentals (35%)	In stable markets, valuations matter most
Normal market	15 – 25	Technicals (35%)	Price action leads in steady conditions
Elevated volatility	25 – 35	News & Technicals (35% each)	Macro events are driving the moves
Crisis / extreme fear	Above 35	News (45%)	In a crisis, macro narrative dominates everything

The PM Agent also runs a divergence check. If the Fundamental Agent is strongly bullish but the News Agent is flagging an upcoming earnings risk, that conflict is surfaced rather than buried. When all four agents agree, confidence is high. When they disagree, the system tells you why rather than just averaging the disagreement away.

When all four agents agree, confidence is high. When they disagree, the system tells you why — and lets you decide what to do with that information.

Before Any Idea Reaches You, It Has to Pass Eight Hard Checks

Once the four agents have scored a stock and the PM Agent has combined their signals, most stock research tools would show you the result. Quant Trading OS doesn’t, not yet. Every candidate first has to pass through a deterministic Risk Engine: eight hard checks that run automatically, that no AI output can override, and that all eight must pass before a stock is surfaced to the dashboard.

Think of it as the compliance desk that sits between the research team and the trading floor. On Wall Street, even the best analyst call doesn’t automatically become a trade, it has to clear risk management first. This is that layer.

Check	What it tests	Why it matters
Data freshness	Is the price data less than 4 hours old?	Stale data produces false signals
Bid/ask spread	Is the spread below 0.5%?	Wide spreads eat your returns before you start
Relative volume	Is today’s volume above the minimum threshold?	Thin volume means the price can move sharply against you
Earnings proximity	Is an earnings report more than 7 days away?	Trading into earnings is a gamble, not a setup
Macro event risk	Is a Fed or RBA decision more than 48 hours away?	Rate decisions can override any technical setup
Daily loss limit	Is the day’s P&L above the configured limit?	Prevents revenge trading after a bad day
Weekly loss limit	Is the week’s P&L above the configured limit?	Protects capital during sustained drawdowns
Signal confidence	Does the composite score meet the minimum threshold?	Weak signals aren’t worth the risk

All eight are configured from the Settings screen and stored in the database — you can tighten or loosen thresholds to suit your own risk tolerance. But the gate itself cannot be switched off, and the AI cannot talk its way past it. If any single check fails, the stock doesn’t appear in the scanner results. No exceptions.

The AI can recommend. The risk engine must approve. You make the final call.

---

What Claude Actually Writes — The Research Brief

Once a stock clears all eight checks, Claude AI receives a structured data package, the composite signal, all four agent scores with their reasoning, the current price and technical indicators, the latest fundamentals, five recent news headlines with sentiment scores, insider activity from the past 30 days, and the current market regime. From all of that, it writes a research brief. The brief always contains the same six things:

• A one-line verdict — the clearest possible summary of the setup (e.g. “Cautious long bias — wait for VWAP reclaim confirmation”)
• Market summary — two to three sentences on what the broader market is doing and how it affects this stock today
• Setup assessment — what the combined technical and fundamental picture shows
• Key price levels — support, resistance, and today’s VWAP, cited from the actual data
• Bull case vs bear case — three specific reasons the trade could work, and three specific reasons it might not
• Suggested position size — expressed in plain English based on the Kelly calculation and your capital band (e.g. “Small — under 1% of capital given elevated VIX”)

The brief is not a buy signal. It is a starting point for your own thinking, the same quality of structured analysis a Wall Street desk would produce, delivered in the time it takes to make a coffee.

---

Does It Actually Work? The Methodology’s Top Picks With Real Numbers

There’s an obvious question that any stock research tool has to answer: does the methodology actually surface good ideas? To answer that honestly, I built an Examples tab into the app, a showcase that runs the real four-agent methodology across the full universe, takes the highest-conviction ideas it surfaces today, and shows how a hypothetical $10,000 position in each would have performed over the past year using real historical prices from EODHD. Before sharing the numbers, three things worth being upfront about:

• These are real historical prices — verified against EODHD’s adjusted close data, which accounts for dividends and splits. No numbers were made up.
• The methodology didn’t fire a buy signal on these stocks a year ago — it rates them highly now, and these are the stocks it currently likes that also happened to perform well over the past year. That’s different from a live backtest, and the app says so clearly.
• Past performance does not indicate future results. This is a research and decision-support tool, not a crystal ball.

With that said, here’s what the snapshot looked like as of 5 June 2026:

Stock	Market	Hypothetical $10,000 position	1-year return
GOOGL — Alphabet	S&P 500	$10,000 → $22,100	+121.0%
AVGO — Broadcom	S&P 500	$10,000 → $16,980	+69.8%
NVDA — Nvidia	S&P 500	$10,000 → $15,940	+59.4%
TLS — Telstra	ASX 200	$10,000 → $10,730	+7.3%
COL — Coles Group	ASX 200	$10,000 → $10,510	+5.1%
Blended (all five)	—	$50,000 → $76,252	+52.5%

The US picks are doing the heavy lifting here , which makes sense given that the methodology’s four agents rate US mega-cap tech names highly on fundamentals, technicals, and smart money flow simultaneously. The ASX names show more modest returns, which also reflects reality: ASX large-caps tend to be slower-moving, dividend-paying businesses rather than high-growth compounders. The methodology scores them accordingly.

One decision I made deliberately: the Examples tab is a frozen snapshot, not a live feed. The numbers you see in the app are the same numbers in this post — they don’t change every time the page loads. That matters for a blog post and a video, where readers and viewers should be able to verify the numbers they’re seeing are the same ones being discussed. If I had used live data, the returns would drift daily and the post would become misleading within weeks.

The goal was never to claim the AI predicted these moves. It was to show that the methodology consistently surfaces high-quality ideas — and let the data speak for itself.

---

Why the ASX Numbers Are More Modest — And Why That’s Honest

One thing I noticed during the build: the methodology gives US stocks significantly higher conviction scores than ASX stocks right now. The top US names were scoring composite signals of +0.2 to +0.4. The top ASX names were scoring +0.03 to +0.07. That gap is real, and the app doesn’t hide it.

The reason comes down to data coverage. EODHD has deep fundamental, news sentiment, and insider data for S&P 500 names. ASX coverage is thinner — fewer analyst estimates, lighter insider filing data, and the Capitol Trades politician flow data simply doesn’t exist for Australian equities. So the ASX agents run with less information, confidence scores are lower, and the methodology reflects that by assigning smaller position sizes to ASX ideas. That’s not a bug. That’s the system working correctly, surfacing lower conviction where the data is weaker, and letting you factor that into your decision.

---

How It Was Built: 13 Sessions, One Rule: Don’t Move On Until It Works

The entire Quant Trading OS was built using Claude Code, Anthropic’s AI coding tool that runs in your terminal and reads, writes, and edits files directly in your project. The way the build was structured is what made the difference. Every session had one objective, a written list of acceptance criteria, and a strict rule. No session could start until the previous one was confirmed working. No exceptions.

Session	What was built	How you know it worked
1	EODHD data layer — the connection to live market prices, fundamentals, and news	Live price for AAPL and BHP.AU returned correctly; cache confirmed working
2	Alpaca client — US live quotes and paper trade execution connection	Live quote returned for AAPL; historical bars returned correctly
3	Data normaliser — converts all sources into one consistent data object per stock	Same TickerData object returned correctly for a US and an ASX ticker
4	Four analyst agents — Fundamental, Technical, News, Smart Money	All four agents returned a valid signal for NVDA and BHP.AU with no errors
5	PM Agent and Risk Engine — signal convergence and the 8-check eligibility gate	Risk engine returned false when any single check failed; Kelly size capped correctly
6	Claude analyst — the AI research brief writer	Full structured brief returned for three live tickers; requires_human_review always true
7	FastAPI backend — all 18 API endpoints wired and documented	Every endpoint returned correct data; /health showed all 9 modules healthy
8	React frontend — scanner dashboard with ranked results, live prices, regime strip	Scanner loaded, ranked 25 stocks, live prices updating every 45 seconds
9	Stock brief view — full research note with chart, indicators, and agent breakdown	Brief loaded for US and ASX ticker; chart rendered; Kelly sizer calculated correctly
10	7-screen onboarding flow — personalised experience based on your profile	Beginner and experienced paths both completed correctly; settings stored
11	Supplementary signals — Capitol Trades scraper and ASX announcement feed	Real politician trades scraped and stored; NVDA insider data returned
12	Trade journal and settings screen — log trades, adjust risk parameters live	Journal entry persisted across full page reload; risk config saved to database
13	Examples showcase tab, per-agent mini signal bars, and US price data fix	5 cards loaded with real returns; 4 mini-bars per scanner row; Today % showing correct single-day moves

By the end of Session 13, the test suite was sitting at 224 passing tests, 18 REST endpoints, and a verified build confirmed through headless Chromium browser automation. The acceptance-criteria-first approach is the reason the codebase is in that shape, not because the AI wrote perfect code every time, but because every imperfect piece was caught and fixed.

If you use AI to build software, the single most important discipline is knowing what “done” means before you start. Without that, you end up with code that looks finished and isn’t.

---

What Phase 1 Is — And What Comes Next

What’s described in this post is Phase 1: a research and decision-support tool. It scans, it analyses, it writes briefs, and it surfaces ranked ideas. It does not place trades. That is intentional. The roadmap has four more phases, each one unlocked only after the previous one is proven:

Phase	What it adds	Status
Phase 1 — Research OS	Scanner, stock brief, four agents, Claude analysis, trade journal. No trading.	✅ Complete
Phase 2 — Backtesting	Run the methodology against historical data to validate whether the signals actually produce positive expectancy before any real capital is involved.	Next
Phase 3 — Paper Trading	Connect to Alpaca’s paper account. Human approval required for every order. Minimum 6 weeks of simulation before advancing.	Future
Phase 4 — Live Trading (US)	Small position sizes, live US equities via Alpaca. Human approval always required. No automation.	Future
Phase 5 — Limited Automation	Only after Phase 4 has been validated over at least 6 months of live trading with consistent results.	Future

The sequencing is deliberate. Backtesting before paper trading. Paper trading before live capital. Live capital before any automation. Human approval is a permanent fixture at every stage that involves real money.

The core idea behind all of it stays the same throughout: AI does the research, the risk engine approves the setup, and the human makes the decision. That’s the only version of this that makes sense to build.

---

Download the Architecture Diagram — Free

I’ve put together a one-page architecture diagram showing exactly how all seven layers of the Quant Trading OS fit together from the data sources, the four analyst agents, the PM Agent, the Risk Engine, the Claude brief, and the dashboard. It’s the same diagram referenced throughout this post, formatted as a clean one-page PDF you can save and refer back to.

No email required. Click to download. A full video walkthrough is coming up soon on the My MBA Project YouTube channel.

---

What Would You Build Next?

The four-agent structure in this build isn’t specific to stocks. The same pattern, multiple specialist AI agents running in parallel, signals combined by a portfolio manager layer, ideas gated by a deterministic risk engine could be applied to any decision that benefits from looking at a problem from more than one angle at once. Credit analysis. Supplier evaluation. Hiring decisions. Market entry assessments.

If you’re thinking about how to apply this kind of multi-agent approach to your own work, drop a comment below or reach out via YouTube. I read everything.

---

---

Disclaimer: Nothing in this post constitutes financial advice. The Quant Trading OS is a personal research and decision-support tool built for educational purposes. Past performance of any stocks mentioned does not indicate future results. Always do your own research and consult a licensed financial adviser before making investment decisions.