Sign in Open account
Platform
Markets Platform Trading & Execution AI Intelligence
Products
Marketplace Desktop Agent API
Knowledge
Research Academy
Company
Enterprise About Trust & Security Legal
Home/AI Intelligence

A reasoning layer with a fence around it

The AI Orchestrator routes tasks to specialist agents and enforces tool permissions and context boundaries. Structured outputs are JSON-schema validated; malformed or unsupported actions are rejected outright.

WatchlistINDICATIVE
SYMLASTCHG
XAUUSDGold / USD2,341.8+0.62%
BTCUSDBitcoin / USD67,412-1.04%
SPX500S&P 5005,412.3+0.31%
USOILWTI Crude78.94+0.19%
Specialist agents

Seven agents, each with a narrow mandate

A single general-purpose assistant would be easier to build and much harder to trust. Narrow mandates mean narrow blast radius when something goes wrong.

01

Market Agent

Turns normalised data into a factual market-state description before any interpretation is offered.

02

Technical Agent

Applies the platform's own indicator and structure methodology, with version IDs attached to every conclusion.

03

News / Macro Agent

Reads external sources as untrusted input. Prompt-injection defences active on every document and web page.

04

Research Agent

Retrieves from the approved knowledge base only, preserving source references for citation.

05

Risk Agent

Explains why a risk decision was made — it does not make the decision. The deterministic gate owns that.

06

Strategy Agent

Describes strategy behaviour and edge assumptions. Output is advisory text, never executable code.

07

Mentor / Journal Agent

Reviews your journal and decision history to surface recurring behavioural patterns.

You are the last agent

Every output from the seven above is a proposal for a human to accept, modify or discard.

Guardrails

What the model is not allowed to do

These are enforced in the orchestrator and the service layer, not written into a prompt and hoped for. A guardrail that lives only in instructions is a suggestion.

  • Model text can never become executable code
  • Model text can never become an order instruction without validation
  • No bypass of the Risk Engine, under any framing
  • All external content treated as untrusted input
  • Allowlisted tools only, each with its own schema
  • Never imply guaranteed returns

Every analysis carries its provenance

Timestamp2026-09-14T07:41Z
Data snapshotsource, version, time range
Model / providerrecorded per request
Strategy / rule versionv2.1.0
Confidence / uncertaintystated where applicable
Why provenance matters: an AI conclusion without a timestamp, a data version and a model version is not analysis — it is an opinion with a nice layout. We show all three, every time.
Knowledge engine

Retrieval from approved knowledge only

Documents carry source, owner, version, approval state, effective date, tags, asset class and an access policy. Only approved knowledge is eligible for production retrieval — draft material cannot leak into an answer.

01

Ingest with metadata

Source, owner, version, approval state, effective date, tags and asset class are attached at ingestion, not added later by guesswork.

02

Chunk and embed, version-aware

Retrieval respects version boundaries, so an answer never blends a superseded rule with the current one.

03

ACL on proprietary methodology

The part of the knowledge base that constitutes trading IP is access-controlled and never surfaces in a public or lower-tier answer.

04

Cite, or say nothing

Answers preserve source references internally and expose citations where the product experience requires them. Hallucination flags are raised rather than hidden.

Get started

Interrogate the reasoning, not just the conclusion

AFH Score, Market Radar and the AI Scanner arrive in Phase 2, each carrying its data snapshot and model version.