US equities · automated execution
Scalping, rebound swings and long-term positions — scanned, entered, hedged and exited automatically, routed to your own Moomoo account. Sign in with your phone number. Everything behind that tap is already built.
The trading clock
Bounce-Off-Band · 15-minute trigger
Three horizons, one account
Each mode runs its own entry logic, its own exit rules and its own risk ceiling. Run one, or run all three side by side.
The Fleet Trades engine. Surge detection fires inside the dip and surge windows, with separate win-exit and loss-exit rules on every position.
Buys weakness that still has structure underneath it. Multi-timeframe momentum has to agree before size goes on.
Broad US index and large-cap exposure, accumulated on a schedule. No discretionary override, no reaction to headlines.
The signal, in full
Bounce-Off-Band · 15-minute trigger
A symbol dumps on enormous 15-minute volume, down into its 4-hour lower Bollinger band, and then prints a green 15-minute candle right there. That green candle's close is the alert. Three conditions, checked in order — all three, or nothing fires.
BOB-TF 15m in the moomoo app — SOXL and its inverse SOXS, 5 August 2026. Band envelopes, volume and reversal markers on the timeframe the trigger fires on.
Scroll the chart sideways →
A 15-minute bucket prints at least 80× the symbol's own baseline — the median of its 96 most recent 15-minute bars. Relative to that symbol, never an absolute threshold, so a thin name and a mega-cap are judged on the same footing. Must have occurred within the last ~2.5 hours.
Bollinger BB(20,2) on the 4-hour timeframe: lower band = mean of the last twenty 4h closes, minus two standard deviations. The spike must close at, below, or within 2% above that band. Measured on closes, not on a wick piercing the band.
The latest closed 15-minute bar must be green — close above open — and still inside the band zone. That close is the alert and the entry. Because it reads only closed bars, a trigger never repaints.
All three conditions met. The green candle has closed at the band — this is the entry signal the live sleeve acts on.
Spike-at-band holds, but no green candle has printed yet. The symbol sits on the extended watch list until one does — or until the setup decays.
Getting started
Enter your number with country code. First-time users get a one-time code over WhatsApp. No email, no password reset loops, no onboarding forms.
OpenD bridges the fleet to your own brokerage account. Orders route to your account and your funds stay in it — the agent places trades, it never holds cash.
Scanning, entries, hedges and exits run on their own schedule. Watch open positions live in Fleet Trades, or close the app and check back later.
Under the tap
Multiple model families review each setup independently. A deterministic risk engine holds the final veto — no model can size a position on its own.
Open-weight models run the continuous scan on your own cluster. Hosted models are called only for narrative and filings work. Disagreement reduces size rather than blocking the trade.
Deterministic indicators generate the actual triggers. Same input, same output, every time — fully reproducible.
Hard limits on per-trade and portfolio exposure. Loss exits are pre-committed before entry, not decided in the moment.
Orders route through the OpenD gateway into your own Moomoo account. The desktop bundle keeps the session live on your Mac.
On-premise inference
Cloud APIs put your signal logic, your positions and your timing on someone else's server, behind someone else's rate limit. Fleet inference runs on NVIDIA DGX Spark nodes you own — eight of them, linked through a 400 Gb/s switch into a single memory pool.
Inference sits beside the execution gateway. No round trip across the public internet inside a two-hour scan window.
The scan layer never leaves the premises. Continuous signal logic, positions and triggers stay on hardware you control.
No rate limits, no provider outage. Throughput during a surge window is yours alone, not shared with a queue.
The model watches, it doesn't wait. Local cost per token is electricity, so inference runs continuously against live tape instead of only on request.
Per-node specifications are NVIDIA's published figures for DGX Spark. NVIDIA documents inference on models up to 200B parameters on a single unit, and up to 405B parameters across two units linked over ConnectX-7. An eight-node deployment is a custom cluster configuration: the cluster column states aggregate hardware totals, and usable model size and token throughput depend on quantization, model architecture and interconnect topology. Hardware shown is a scale drawing, not a product photograph.
What we're aiming at
Start on your phone