> For the complete documentation index, see [llms.txt](https://docs.visionmakers.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.visionmakers.io/whitepaper/architecture.md).

# 5 · Architecture

Four layers from raw chain state to a signed transaction

> **FIG-03**: Ingestion to interface, with one intelligence core shared by every product.

## Layer 1 · Ingestion

Parallel pipelines bring in blockchain node state across EVM networks and TON, decentralised and
centralised market data, social graph activity, news feeds, and behavioural telemetry from the
platform itself.

Every record carries **lineage metadata**: source, observation time, and the transformation chain
that produced it. This is not bookkeeping for its own sake. A system whose product is
interpretation must be able to show its work, and lineage is what makes an explanation auditable
rather than merely confident. It is also what makes compliance reporting and incident
reconstruction tractable.

## Layer 2 · Intelligence

Raw events consolidate into an event lake, then pass through feature engineering and three model
families:

- **Scoring**: influence, adoption and liquidity health rendered as comparable numbers rather than raw counts.
- **Graph models**: relationships between wallets, tokens and communities, which is what makes concentration and coordinated behaviour visible.
- **Anomaly detection**: deviation measured against a token's own established pattern rather than a global threshold that suits nothing in particular.

Signals are **source-credibility weighted**: a social account's contribution is scaled by its
historical accuracy, not its follower count.

## Layer 3 · Reasoning

Four jobs sit here.

**Semantic resolution** turns language into a structured query. *"What's trending right now"*
becomes a ranked mindshare query with an implied window; *"is this safe"* becomes a composite over
liquidity, concentration and contract flags.

**Model routing** selects the model appropriate to the request. Cheap classification work does not
need a frontier model; multi-step reasoning over conflicting evidence does. Routing is a cost and
latency decision made per request.

**Agent orchestration** runs multi-step workflows, scheduled briefs, conditional monitoring,
pre-staged execution, with every state transition logged against the user's account.

**Adaptive explanation** calibrates output to the reader. The same query returns a plain
explanation with defined terms for a new user and a dense metric summary for a professional. The
computation is identical; only the presentation differs.

## Layer 4 · Interface

Web platform, Telegram AI Hub, Telegram trading bot, mobile applications and a developer API,
with role-based access control running from free tier to institutional.

<div class="callout">
<span class="callout-label">Architectural constraint</span>
The reasoning layer has <strong>no privileged execution path</strong>. Any action it can prepare
is one the user could take through the graphical interface, under identical permission checks,
and every one terminates at a signature request. The conversation is a faster route to the same
controls, never a wider one.
</div>

## Why layering pays

Because intelligence is shared, a new product does not need its own data team. vDEX, MOFO and
vDEX and the Intelligence Layer all consume the same liquidity and sentiment features. The
marginal cost of the tenth product is therefore far below the cost of the first, the structural reason an
nine-product surface is viable for a company of this size.