DD AI & Automations← Back to the board
Talk Systems
Buzz watches the markets around the clock, decides when to get in and out, and places every order on its own. Nobody sits at a screen.
Eight separate strategies run at the same time across two accounts, and everything they want to do has to pass a risk layer that can overrule them.
It took years of nights and weekends, and most of what I built got deleted along the way.
The two accounts do different jobs, and a strategy automatically picks up the rules of whichever one it runs on.
Each strategy is a self-contained method with its own entry, exit and risk settings. They deliberately cover different markets so they don't all win or lose together.
The part that works out what to trade is deliberately separate from the part that places the order, so a fault in one can't quietly corrupt the other.
Every account keeps an explicit list of which strategies are allowed to open a position. Anything not on it gets refused by name — including a strategy added later, which is exactly when a permissive default would do damage.
As losses deepen, the system cuts its own position sizes in stages, and eventually stops opening anything at all until things recover.
Some of the worst outcomes come from trading at the wrong moment rather than picking the wrong direction. These rules cover the moments.
When a broker replies "ok", that is not evidence the order actually went through the way it was meant to. Every position gets checked against the broker's own records instead of trusting the reply.
A program can be running and still be doing nothing useful. This layer checks each part is genuinely alive by looking at whether it's still producing fresh output.
That rule exists because a monitor once compared the wrong account's data and blocked a restart hundreds of times across fifteen hours without raising an alarm.
Nothing goes live on a backtest alone. A strategy has to survive several rounds of testing and a spell on simulated money first — and it keeps being re-tested afterwards.
Everything reports into one place, and any part of it can be stopped from a phone.
Funded trading accounts come with limits that end the account if you break one even once — a maximum you can lose in a day, a maximum overall, a minimum number of days traded, and restrictions around weekends and news.
Almost all of the risk design above exists to respect those limits automatically instead of depending on someone watching a screen. It behaves differently during an evaluation than once funded, and the same design works just as well on your own money.
Eight automated sleeves across two account brains, designed to run unattended. An engineering deep-dive, not a performance page.
One for intraday and event trades, one for multi-day positions, with different time, weekend and news rules.
Default deny. An allowlist per account decides which strategies may open a position; the gate never blocks exits.
Risk steps down automatically as drawdown deepens, then stops entirely. Anchored to the algorithm's own curve.
An acknowledgement is not an execution. Stops are verified inline and swept every 180 seconds, and positions reconciled against the broker.
Heartbeats measured by log freshness rather than whether a process exists.
Multi-tier validation, paper and shadow bridges, walk-forward revalidation and decay alerts. Nothing gets armed on a backtest alone; failures are documented and killed.
Designed to respect prop firm account rules automatically — daily loss limits, maximum drawdown, minimum days, weekend and news constraints — and the same design applies to your own capital.