Solutions · Hedge Funds

The AI trading stack your ODD questionnaire was written for.

Every fund deploying AI in the investment process now faces the same allocator question: how do you supervise it? Most answers are policy documents. VegaTrade's answer is architecture — the platform was built so that an unsupervised, unexplained, or unreproducible AI decision is not a violation that gets caught, but a state that cannot occur.

Glass-Box AI

Show the investment committee what the model knew.

Every AI inference on the platform emits a paired record into the audit chain: the prompt context, model identifier, and an SHA-256 hash of canonical inputs at start; the output, the reasoning, and a reproducibility-class discriminator at completion. The decision your portfolio manager sees in the console is the same record a forensic reviewer reconstructs years later — byte-identical, from immutable storage.

That changes the texture of an ODD conversation. "What was the model's reasoning on the day it cut exposure?" stops being an essay question and becomes a database query.

Glass-Box AI Decision Lifecycle
Glass-Box AI Decision Lifecycle INFERENCE STARTED Captured before the answer prompt context model identifier + temperature SHA-256 of canonical inputs tenant + decision identity MODEL REASONS Slow loop only regimes · signals · sizing seconds-to-minutes cadence never on the order path multi-provider: Anthropic · Gemini INFERENCE COMPLETED Captured with the answer output + reasoning seed or provider sentinel reproducibility-class discriminator confidence + sizing rationale PER-TENANT AUDIT CHAIN · DEDICATED AI-DECISION EVENT RANGE Replayable years later: what the model knew, what it decided, and why Black-box AI is forbidden by architecture, not by policy. If an event fails to capture this payload, it is quarantined — never silently dropped.

Strategy Validation

Backtests that charge themselves the full cost of being wrong.

The platform's validation gates encode the anti-overfitting literature as architectural preconditions: walk-forward cross-validation, nested inner cross-validation for hyperparameter selection without peeking, deflated Sharpe ratio to correct for multiple testing, and probability-of-backtest-overfitting. No strategy reaches paper trading without clearing all of them, and the gate verdicts are themselves audit events.

Execution-model fidelity bands quantify what the backtest charges for fills, and once a strategy is live, a continuous backtest-to-live divergence metric tracks whether production behavior is drifting from what was validated — with sustained breach triggering retirement review. Crowding and regime shift surface as quantified deltas, not as quarter-end surprises.

Publishing discipline matches: the platform reports net-of-all-fees Sharpe alongside gross, and after-tax alongside pre-tax for tax-aware strategies. The number that survives diligence is the number the platform leads with.

Supervision

The AI proposes. Your operators approve.

AI-authored strategies are bounded parameterizations inside a constrained grammar — never free-form code — and they enter an approval queue where a human approves or rejects, with both the proposal and the disposition chained. News-driven analysis is propose-only by construction. Per-symbol decisions execute inside approved bounds with per-decision reasoning traces.

The division of authority is the one your compliance manual already wants: models recommend, humans authorize, the chain remembers.

VegaTrade AI Decisions table showing per-symbol buy, sell, and hold decisions with the exact model identifier, reproducibility-class badge, confidence, sizing percentage, and a reasoning-trace button on every row
The AI decision blotter: model identity, reproducibility class, confidence, sizing, and a reasoning trace on every row. Simulated paper-trading environment with illustrative data. Not real market executions, not real funds, and not a representation of any trading results.

Operations

Risk gates that cannot be routed around.

Every order passes a per-tenant risk gateway in the C++ hot path — position-size caps, daily-drawdown stops, order-rate limits, symbol whitelists, and the global kill switch. The gateway is the sole path to a broker; there is no privileged route past it for an operator, an administrator, or a model. Tenants are isolated at the database boundary, and your fund's positions, orders, and strategies are never visible to another tenant by construction.

The platform never custodies fund assets — execution routes to your own broker relationships, and the platform's role is decision, control, and evidence.

VegaTrade dashboard showing buying power, open positions, P&L, equity curve, risk-limit utilization meters, and kill-switch state, with a banner stating the environment is simulated paper trading
Risk-limit utilization and kill-switch state are first-screen dashboard surfaces, not buried admin panels. Simulated paper-trading environment with illustrative data. Not real market executions, not real funds, and not a representation of any trading results.

Bring your hardest diligence questions.

The platform is pre-launch and onboarding design partners. The procurement conversation — architecture walk-through, audit-chain demonstration, evidence mapping — starts at the contact form.

Request Access