I Copied an AI Portfolio Manager's Whitepaper and Built My Own — Here's How

An AI firm built a portfolio that returned +76.4%. Then they published exactly how they did it. Here's what happened when I actually read it.

A few months ago, I came across something I couldn't stop thinking about.

An AI portfolio management service called Autopilot had published verified, live returns on two of their AI-managed portfolios: +76.4% all time on the GPT Portfolio, +61.2% on the DeepSeek Portfolio. 

That alone would have been enough to catch my attention. But then I found something else: the research firm behind the methodology — Dr. Lira's AI Finance Labs — had published a full whitepaper detailing exactly how the GPT Portfolio works. Every prompt, word for word. Every data source. The exact 97 financial variables scored per stock. The complete scoring hierarchy from macro to sector to company to news.

The blueprint was sitting in plain sight. And almost nobody had done anything with it.

That's when I started building.

What You're Actually Paying For

Autopilot is a subscription service that lets you mirror AI-managed portfolios on your own brokerage account – including Robinhood, Charles Schwab, and E-Trade. It's a clean product. But when I looked at what was actually under the hood, something became clear: you're paying for someone else's implementation of models and methodology that are both publicly available.

The models — Claude, GPT, DeepSeek — cost $4 a month in API tokens to access directly. The methodology is documented in a public whitepaper. What the subscription buys you is someone else running it on your behalf, on their infrastructure, improving their system with every dollar you pay them.

I've spent my career looking for situations where you can acquire the asset instead of renting access to it. This felt like one of those situations.

A traditional active manager charges 1–2% AUM. On $100K, that's $1,000 to $2,000 a year, whether the portfolio performs or not. Autopilot's subscription is lower than that — but it's still a recurring fee for a system you'll never own. The number I kept coming back to: what I built runs for roughly $48 a year. Full ownership. Every improvement permanently mine.

Building CIA — Claude Investment Advisor

I named the project CIA. Claude Investment Advisor. The goal was straightforward: replicate the GPT Portfolio methodology end-to-end using Claude, fully automated, running on my own machine.

The core pipeline came together in phases.

First, the data layer. A Python script pulls financial data for all 503 S&P 500 stocks via yfinance — free. A second script pulls news, SEC filings, earnings context, and macro data — also free. From there, Claude Sonnet scores every stock on a scale of 1 to 100 using the whitepaper's exact prompt structure, evaluating all 97 financial variables per company. That scoring run costs roughly $3. Claude Opus then builds a 15-asset weighted portfolio from the top-ranked stocks — another $0.50. The whole pipeline executes trades on a $100K paper trading account through Alpaca.

The first full run completed June 4, 2026.

Phase two was making it self-running. I built an SQLite database to track every position, score, rebalance snapshot, and investment thesis. Then I migrated from a basic scheduler to macOS launchd — a native macOS automation tool — so the system runs through sleep and wake cycles without any manual intervention. CIA now operates on three automated cadences: a full monthly pipeline, a weekly re-score of held positions only, and daily SEC monitoring. I touch it once a month to review. Never to operate.

Closing the Gaps — and Going Further

Phase three was about closing the distance between what I'd built and what the whitepaper specified.

I expanded from roughly 25 financial fields to the full 97 the whitepaper calls for: valuation multiples, margin profiles, balance sheet metrics, short interest, governance risk scores, analyst price targets. I added a macro forecasting module — Claude Opus synthesizes raw market data into a structured three-month forward outlook before any stock is ever scored, matching the whitepaper precisely. I built a sector scoring layer: 11 separate Claude calls, one per GICS sector, establishing the hierarchy the whitepaper specifies — macro first, then sector, then company, then news.

Then I built something the whitepaper doesn't have.

Every weekday at 7am, CIA checks SEC EDGAR for new 8-K filings on every position it holds. If a filing is found, Claude reads it, compares it against the original investment thesis stored in the database, and returns a severity rating — low, medium, or high — with a one-paragraph explanation of what changed and why it matters. High severity flags the position immediately and marks it as thesis-risk in the database.

The whitepaper doesn't do this. A subscription to Autopilot doesn't do this. A system you own can.

"The edge was never the secret sauce. The edge is that almost nobody is willing to flip over the stones that are already sitting right in front of them.”

Where CIA Stands Today

CIA is currently running on paper trading — $100K simulated, not live capital. The +76.4% and +61.2% figures are from Autopilot's live portfolios, not CIA's own track record. I want to be direct about that. The goal isn't to claim those returns. The goal is to own a system built to the same specification — and to keep improving it in ways a subscription never could.

The next milestone is moving to live execution once paper performance is established, and then tracking CIA's returns directly against the Autopilot benchmarks. That's when the comparison gets interesting.

The Bigger Point

We are at the earliest stage of an era where a single person with Claude Code and a few hundred dollars a year can build systems that used to require entire teams. The GPT Portfolio whitepaper was sitting there for anyone to read. Almost nobody will build anything from it.

That gap — between the people who use AI tools and the people who build with them — is going to be one of the defining economic divides of the next decade. And it compounds. The person who starts building today has a five-year head start on the person who waits until it feels obvious. In five years, that head start shows up as systems that are five years smarter, five years more refined, five years more integrated into how you operate. The renters will still be paying for access to someone else's version.

You don't need to be a developer. You need to be curious enough to ask: what if I just tried to build this? Claude Code handles the rest.

CIA isn't finished. That's exactly the point. The value isn't in shipping something final — it's in owning a system that compounds with you, not for someone else.

The stones are there. Most people just don’t put in the effort to flip them.