Vhalanx runs a book of specialist trading agents across day-trade and swing horizons. Every order they propose is judged in real time by a reinforcement-learning governor before it ever reaches a broker — and the same engine is multi-tenant, so anyone can run their own agents on it.
↑ A live simulation of the decision loop. Agent proposals stream in; the governor scales, throttles, or halts each one.
Most trading bots are a single strategy with a fixed rule book. Vhalanx separates finding edge from managing risk — and puts a trained model in charge of the second job.
Each agent does one thing well and stays honest about its edge. No single strategy carries the whole book, so no single failure sinks it.
A PPO-trained governor sits between every agent and the broker. It reads regime and exposure and scales each order — allow, throttle, or halt — the way a human risk desk would.
Sizing, P&L, and stops read live broker state, not internal tallies. What the account actually holds is the source of truth end to end.
The same architecture runs fast intraday strategies and multi-day swing positions — the horizon changes, the governance doesn't.
Independent of any agent's opinion, the engine enforces its own limits — and an operator can override all of it in one click.
Capital is split into independent sleeves — by broker and by horizon — each with its own risk envelope, its own P&L, and its own record. Nothing is ever commingled.
Endpoint is the actual broker-truth result: +$19K / +11% NetLiq on the paper IBKR swing sleeve since 2026-05-13. The intermediate path is illustrative — a weekly index, not a tick-by-tick reconstruction.
Vhalanx is multi-tenant from the ground up. A new user brings their own account, picks their own universe, and gets the full engine — data feed, agents, governor, and controls — walled off from every other tenant.
A tenant connects an Alpaca or IBKR account and configures the symbols they want to trade. The engine syncs their universe on the next cycle — no redeploy.
Free-tier users share a standardized pool broadcast over Kafka from a single market-data publisher. Paid tenants plug in their own SIP feed. Cost scales with symbols, not users.
Per-user universes, per-user P&L, and per-user state. One tenant's fills, positions, and controls never leak into another's dashboard.
From the dashboard, an operator can pause, flatten, or nuke any sleeve or tenant — the command travels end-to-end to the live engine and takes effect on the next cycle. Loopback-only, defense in depth.
Fills carry a lifecycle ID and parent agent, so per-agent P&L is exact — and the governor's reward signal is de-censored, feeding the next round of training.
The claims below are broker-truth or backtest results — labeled exactly as they stand. Paper is paper; only the ring-fenced sleeve is real money.
A Kafka data spine, a stateless engine that ships by immutable tag, and infrastructure defined as code — so the same system runs one book or many.
Narrow agents find the edge. A learned governor keeps the book alive. Multi-tenant from day one, proven on paper, and already trading real money.