AI · Finance

AI Multi-Asset Trading Workstation

Multi-agent AI analysis behind hard risk gates and a four-tier rollout, connected to the MetaTrader 5 terminal SDK.

LiveAI · FinanceMid-size projectPC · Windows desktop

Not investment advice

Client: Proprietary trading operator (in-house tool)Undisclosed by agreement
Industry: AI · Finance
Platforms: PC · Windows desktop
Scale: Mid-size project · 2026

01Background

The client is a proprietary trading operator running precious metals, energy, crypto and FX CFDs through a broker’s MetaTrader 5 terminal. Watching markets, reading macro data, tuning prompts and deciding whether to trade at all were scattered across the terminal, chat windows and spreadsheets. They needed one workspace where AI could do the analysis, but every order stayed controlled and traceable.

Client: Proprietary trading operator (in-house tool)
Markets: Metals, energy, crypto and FX CFDs
Execution: MT5 terminal on Windows
Deployment: Single Windows host

02Challenges

1Fragmented tools

Positions lived in MT5, analysis in chat and risk in someone’s head — none matched up per trade.

2Ungated AI orders

A model could say “buy” with nothing enforcing size limits, loss streaks or market hours.

3Tangled logic

Entry patterns and macro judgment shared one prompt, so every edit rippled and nothing was testable.

4Manual rollout

Going from paper to live meant trading smaller by hand; tracing a trade meant digging through chats.

Let AI analyze, but keep order flow gated — what was missing was a control desk.

03Solution

Shanku split the workspace into three tiers: a PC console, a business core and a bridge sitting next to the MT5 terminal. Rule-based strategies decide when to enter and exit, AI decides whether and how much to trade, and every order passes one hard risk gateway before a rollout tier decides whether it is logged, simulated or sent to the terminal.

Multi-agent analysis

Reflection, technical, fundamental and risk agents report in; only the orchestrator can trade, and it must log its reasoning.

Hard risk gates

Lot caps, one position per symbol, concurrency, daily loss, loss streaks, rate and spread checks live in code, not prompts.

Four-tier rollout

Shadow logs only, paper fills simulated, micro trades minimum size, live uses full size; one account-wide tier, shadow by default.

Layered gold logic

AI writes a pre-session market regime snapshot, rules enter on closed-bar pullbacks, and a position agent manages stops and exits.

MT5 SDK bridge

A local Windows bridge is the only process touching the terminal; if it drops, the dashboard degrades instead of failing.

Prompts and keys in-app

Agent prompts are edited in the console and apply on the next run; API keys sit in a vault and show masked.

04Core workflow

How an order intent gets cleared

1
Engine
Symbol enabled, engine running
2
Market
No weekends, rollovers or news
3
Intent
Rule signal or orchestrator call
4
Risk
Size, positions, loss, frequency
5
Tier
Log, simulate or route to MT5
6
Execute
Micro or live via the MT5 bridge
7
Audit
Events and trades share a run ID

Hard gates exist only in code and AI tools cannot reach the terminal directly; stopping an engine halts all automated analysis and orders for that symbol.

05Results

One write path

However many agents analyze, opens, modifications and closes all pass a single auditable gateway.

Reversible rollout

Shadow through live share one path, so integration and channel checks need no real capital first.

Traceable trades

Positions, fills and events link to an AI run ID, opening the exact analysis behind each one.

Testable rules

Entry patterns are plain rule functions and macro views are expiring snapshots, each verified on its own.

06Screens

Screens are prototypes or rebuilt with sample data. They contain no real client data, amounts or accounts.

07Tech stack

Backend: FastAPISQLAlchemy 2AlembicAPScheduler
AI orchestration: LangGraphDeepSeekOpenAIAnthropic
Frontend: Vue 3Element Plus
Data & integrations: PostgreSQLRedisMetaTrader 5 SDKFRED macro data
Deployment: Windows hostDocker (database & cache)

This case study demonstrates software development capability only and does not constitute investment advice. Screens are illustrative and contain no real amounts, positions or accounts.