Members · 02
The question I get most often: what does this actually run on? Here is the architecture, layer by layer. The engine, the data, why backtests run in two places, where the AI sits, and where live trading happens. Every piece of it is available to you. Nothing here needs a fund's budget.
Two things are worth seeing before the detail. First, there are two ways in for you, the human: you talk to the AI, or you look at the console, the dashboard we built for visibility. Second, the AI is not one box in a chain. It can see and operate everything on your computer and in the cloud.
We chose QuantConnect as the engine and the provider, for two reasons. The first is LEAN, their open source trading engine. The same engine runs a backtest and runs a live account, so a strategy is written once and never translated between research and production. The second reason is everything around the engine. QuantConnect also supplies market data, cloud machines to run backtests on, and live trading servers that sit next to the main US markets. Building that ourselves would be a business in its own right. Renting it means we only build the parts that are actually ours.
LEAN understands two languages, Python and C#. We use Python. All of our code lives on our own computer as ordinary files, and we work on it there. The bridge to QuantConnect is the LEAN CLI, a command line tool: you type a command such as "run this backtest in the cloud" or "deploy this system live" and the CLI sends it to QuantConnect and brings the results back. Because everything is a plain typed command, it is also exactly the kind of thing an AI can drive, which is where the next section comes in.
Our build agent is Claude Code. Rather than living inside one step, it sits over the whole map. It edits the files, runs the tests, sends backtests to the cloud or runs them locally, reads the results out of the database, and deploys systems when we say so. In practice, most of our day is a conversation: describe the system we want, the AI builds it on the shared foundation, we look at the results together, and we decide what to try next.
One rule keeps this sane. The AI builds, the gauntlet judges. Nothing is trusted because the AI wrote it. Every result is recorded and has to survive the same testing as anything written by hand.
A backtest can run in two places. In the cloud, it runs on a QuantConnect backtest node, a rented machine that sits beside their market data. That is where our US stock and futures tests go, because those need decades of clean history that still includes the companies that went bust, and QuantConnect already holds it.
Locally, it runs on our own machine inside Docker. Docker is a tool that runs a program in a sealed box with everything it needs already inside, so the LEAN engine runs identically on any computer without a day of installation. Local backtests are optional. You could run everything in the cloud and be fine. We keep the local option for three reasons: it is free and unlimited, so early exploration costs nothing; it never queues behind the one cloud node; and for crypto and FX, where we download the data ourselves, the data is already here. The rule of thumb: backtests run wherever their data lives.
When a system has earned its place, it is deployed to a QuantConnect live node, one per brokerage account, running around the clock and sending orders straight to the broker. Monitoring watches each node's heartbeat and the login clocks that brokers impose. The account level is where the sleeves from the orientation page become real: which systems run in which account is a deliberate mapping, shaped by each account's rules.
Underneath everything sits one results database. Every backtest, local or cloud, and every live run lands there with the exact settings that produced it. No side spreadsheets, no "I think this was the good run". The console, our dashboard, reads from that database and nowhere else, which is how we get a single view of systems, backtests and live deployments. It gets its own piece shortly.
Less than you might think. A computer: laptop or desktop, Windows, Mac or Linux. We use a Mac. An AI agent to build with; ideally one that can control your computer, not just chat. We use Claude, specifically Claude Code, but the approach works with your agent of choice. And a QuantConnect account, which gives you the engine, the data, the cloud backtests and, when you are ready, live trading. The details of each of those get their own short piece next.