For the first stretch of this the whole system ran on a spare laptop on a desk in my dorm room. It worked, in the sense that it was running, and it was never once out of my mind: a closed lid, a dropped connection, a roommate moving it, an update deciding two in the morning was a good time to restart. A trading loop that stops halfway through a position is worse than one that never started. So it moved onto an AWS box — the smallest Lightsail instance there is, in Ohio because that is the region nearest the exchange, with a static IP, the firewall cut back to SSH from my own address, and the runner registered as a systemd service so a reboot brings it back without me. The dashboard is reached over Tailscale rather than through an open port, which means I can look at it from my phone without the internet being able to. What stayed on my laptop is everything heavy: the years of one-minute bars, the backtests, the optimisation runs. Those want a real machine for an hour at a time and they want it for free, and the box only has to hold one live loop and one broker connection, which is why it costs what it costs. I set out to buy peace of mind and came away having learned AWS properly, which was the better half of the trade.

The first version of all of this was TypeScript from end to end, written with another developer I was working with at the time. Then I stepped away from it for a while, and when I came back I was on my own and had to read it as a stranger would. Rewriting the backend in Python was the decision I put off longest and should have made first. It is not that the language is better. It is that every library this problem needs already exists there — pandas and NumPy and Parquet, the broker and data vendor SDKs, the statistics — and, more than that, so does everything anyone has written about the problem. Every paper, every notebook, every answer to the exact question I was stuck on, all in Python. The interface stayed TypeScript, where it is the right tool and where React and a charting library do in an afternoon what would otherwise take a week. What the split bought was speed: an idea now goes from a sentence to a tested result in a fraction of what it used to take, and the honest part of that figure is not the language, it is that rewriting the thing alone was how I finally understood every piece of it.

The system runs on a machine I never look at, which is the whole point of the machine and also the problem with it. A log file you have to go and open is a log file you do not open. So every entry and every exit posts to a Discord channel the moment it happens: the instrument, the side, the size, the price it went in at, and the stop and the target it is going to be judged against — then, at the other end, where it actually closed, in dollars and in points, with the account value and the strategy's running total underneath. The line I read first is the mode, because it says which account this was, a test run or a paper account or a prop firm evaluation, and a message that does not say that is a message you can misread badly. Two rules hold the whole thing up, and both of them are about the trading loop rather than about the messages: the notifier can never raise, because a listener that throws gets dropped and the alerts would go quiet without telling me, and the post is fire and forget, because a webhook having a slow afternoon must not be able to stall a bar. What it changed most is how I find bugs. Exits recorded at the entry price, a position the venue had already closed that the runner still thought was open — those are a fortnight of not noticing in a log, and about four seconds of noticing on a phone.