Hadi SoufiAI Systems Architect

ZVAKTHOR · Agentic trading infrastructure

Separating intelligence from control.

An algorithmic execution platform for global financial markets. Specialist agents analyse markets independently, while an isolated Rust core validates every decision for risk and execution eligibility before any order reaches the market. Every decision — accepted or rejected — is recorded and replayable.

The problem

Most agentic AI stops at a working demo.

The hard part is everything after it: execution governance, observability, and failure handling. A system that can generate a signal is not the same as a system that can be trusted with capital.

The Operating-Layer Dominance Hypothesis (Paper 01) argues that agent behaviour is driven by interface and execution design, not by model reasoning quality alone. That hypothesis is the starting point for the architecture below.

Architecture

Three layers. One enforced boundary.

Control boundary — nothing crosses unvalidated

The split is deliberate: a component that can influence a decision is not the component that authorises it.

What the architecture enforces

Five guarantees.

  • No unvalidated order. Every decision passes risk and eligibility validation in an isolated control core before it can reach the market.
  • Full decision lineage. Every decision — accepted or rejected — is recorded and replayable.
  • Bounded authority. Authority carries explicit scope, validity and revocation, and is revoked at the authorisation boundary before subsequent execution.
  • Reproducibility. A decision is reproducible only from a versioned execution context: software, policy, model, configuration and dependency versions.
  • Failure containment. A failed transition must not expose a valid-looking partial state; committed state survives process or node failure.

Custody and control

Non-custodial by design.

ZVAKTHOR is designed to be licensed to clients who run it on their own broker accounts. Capital stays in the client’s name, and the system never takes discretionary control.

This is an architectural constraint rather than a contractual promise: because the control core validates execution eligibility instead of holding keys or balances, the platform does not need custody to operate.

Status

Core architecture and the first agent group validated end to end in a demo environment. Next phase: full agent activation and live-account validation, with remaining agents activated as their data infrastructure is funded. No live performance figures are published here, and none should be inferred.

Research behind the architecture