MARY

What MARY Is

MARY is an investment reasoning architecture — a system built to analyze the questions that matter most to institutional investors: What is the macro regime? Which stocks are worth owning? How should a portfolio be constructed given current conditions? Which strategies work in this environment? When should you get out?

Every answer is grounded in real data. MARY pulls from 21 FRED macroeconomic indicators, SEC filings, and live market prices, and runs deterministic computations — regime scoring, portfolio optimization, valuation models, backtests — before any language model synthesizes the result. The reasoning is explainable. Every claim traces back to a source.

But the data does not just flow in. It flows back. Every output MARY produces — every regime call, every allocation, every signal — becomes training material for the next cycle. The system does not stay still. It learns from what the market confirms, adjusts from what it gets wrong, and arrives at each new day a little sharper than the last.

Built by a team of financial and AI experts, led by Parson Tang — an investment professional with 20+ years of experience across major financial institutions and global markets. The system reflects how practitioners actually think through investment problems — not how a chatbot summarizes financial news.

How MARY learns from every market cycle →

How the Agents Work Together

Each agent operates independently on its domain. But the most important questions in investing cross domains — and that is where MARY's multi-agent architecture matters.

When a cross-domain question arrives, MARY dispatches to the relevant agents simultaneously. The Macro agent informs the Portfolio agent's position sizing. The Fundamental agent's quality scores inform stock selection within Portfolio. The Alpha Lab's regime-fitness scores gate which strategies to recommend. When all four agents converge on the same conclusion, MARY flags it explicitly.

The synthesis step identifies where agents agree, where they diverge, and which view carries more weight given the confidence levels. The result is not a summary — it is a reasoned, integrated position.

The Intelligence Flywheel

Most AI systems are static — they ship a model and it degrades over time. MARY is designed to compound intelligence. Every signal she generates, every regime she classifies, every trade she recommends feeds back into her learning pipeline.

The result: the longer MARY runs, the sharper she becomes. New data becomes training data. New strategies generate new insights that inform better strategies. This is compounding intelligence — the same principle that makes the best investors better over decades, encoded into a system.

Machine Learning That Earns Its Seat

There is a reason most investment AI disappoints: the machine learning is decorative. A model is trained once on historical data, shipped, and left to slowly drift as markets evolve. The world changes. The model does not. Eventually, the gap between what the model learned and what is actually happening becomes too wide to be useful.

MARY was built differently. Her machine learning layer does not merely consume data — it generates its own. When the macro engine reads 21 economic indicators and produces a regime classification, that classification becomes training data. When the portfolio engine makes an allocation decision and the market grades it over the following quarter, that outcome feeds back in. The system is continuously writing its own curriculum from the markets it is watching.

This creates a capability that static systems simply cannot replicate: the ability to form independent views. MARY does not just consume published macro data — she synthesizes her own macro assessments, independent of consensus. When her indicators pointed to stagflation — rising prices, slowing growth, deteriorating credit conditions — MARY classified the regime as stagflationary before that word had entered the mainstream conversation. Not because someone told her to. Because the pattern was there in the data, and she had learned to recognize it.

This is what self-improving intelligence means in practice. The model that runs today has seen every regime MARY has ever classified, graded against every outcome that followed. It has learned which indicators matter most in which environments, which signals are noise and which are signal, and how the relationships between variables shift as the cycle turns. And it will know more next quarter than it does today.

That is a different kind of edge — one that compounds quietly, every single day.

Built to Be Trusted

Intelligence without accountability is just noise. MARY is designed so that every answer can be interrogated — where the data came from, which model produced it, and under what conditions the view would change. That is not a disclaimer. It is the architecture.

See MARY in Action

The research published on this site is MARY's actual output — regime assessments, company analyses, portfolio frameworks, backtested strategies. Read the work and judge the reasoning.