The account has been running for three months. Four strategies. Real Hyperliquid perpetuals. No human checking charts.
That's the version I wanted to build. Getting there took longer than I thought.
I came at this backwards — started with the system, not the strategy. Most people learn one strategy and try to automate it. I built the infrastructure first, then added strategies on top. Looking back, that was probably right. The infrastructure is the hard part.
The Stack
The system runs on two VPSs. One dedicated to trading — executes fills, manages stops, reports state. One for AI orchestration — runs analysis, signal routing, and the circuit breaker. They talk to each other. Neither is optional.
Four strategies run simultaneously: whale copy (tracking large Hyperliquid positions), vol spike, liquidation cascade, and OI divergence. Each gets equal capital. If one hits -20% loss, it pauses. The others keep running. The portfolio doesn't stop because one leg had a bad week.
The Part I Got Wrong
The circuit breaker took three attempts. First version was too aggressive — it halted after a bad morning, missed the afternoon recovery, and I'd come back to flat. Second version was too permissive — I hit max drawdown twice before it triggered. The current version is Kelly-based: if expected edge goes negative, everything stops. Not triggered by losses. Triggered by whether the math still works.
Stop losses come from three places simultaneously. The bot sets a trailing stop. The exchange holds a backup Stop Market order. Leverage is hard-capped three times below max. Any one of them can close a position. Two running at once is normal.
What the Returns Actually Look Like
The thing I didn't understand at the start: returns from a well-built system are boring. Not in a bad way. Boring because the system does what it's supposed to, consistently. No dramatic reversal. No clutch recovery. It either had edge or it didn't — you find out in the data, not in the moment.
I started writing code two years ago. This is the most technically demanding thing I've built. Also the most honest — financial markets don't care about your architecture choices. The system either works at 3am on a Tuesday or it doesn't.