Introduction
A behind-the-scenes look at how internal FOREX trading teams can use large language models (LLMs) to answer questions about customer onboarding, trades, and currency risk — served in real time using VLLM and AWS.

What Kind of AI Agent Are We Building in This Post?
This blog post covers the foundation of a chat-based AI assistant for internal FOREX traders and operations teams.
-
Built to respond in real time to complex questions
-
Trained to understand user roles and limit access to only their assigned books and customers
-
Designed to extend easily into future voice input or document-based workflows
-
Operates within enterprise-grade security and latency requirements
This is not a general chatbot. It’s a task-specific, scoped, and high-performance agent for decision-makers in FOREX risk management.
Business Context
FOREX trading desks manage fast-moving, high-value currency trades across global markets. Each trader or operations user works on specific books and manages customers with different risk profiles, trade types, and compliance needs.
They need to:
-
Instantly access relevant data
-
Get summarized risk, exposure, and trade answers
-
Ask questions in natural language without learning a new interface
-
Only see the data they’re authorized for
Instead of clicking through dashboards or writing SQL queries, they simply ask:
“What’s my GBP/USD exposure today?”
“Did Cust104 complete onboarding for options?”
“Who added trades over $10M this morning?”
To enable this, we use LLMs deployed with VLLM for high-speed inference and AWS Bedrock for secure enterprise integrations.
Users
Trader
Responsible for managing positions, executing trades, and monitoring book-level performance.
Risk Analyst
Oversees overall exposure, compliance thresholds, market shifts, and forward-looking simulations.
Onboarding and Compliance Ops
Manages KYC, AML, and internal approval workflows for new or updated customers.
Their Needs
Trader Prompts (30+ Examples)
-
What’s my EUR/USD exposure today?
-
List customers trading USD/INR in the last 24 hours.
-
Show all spot trades over $10 million.
-
Did anyone open a GBP forward contract today?
-
Summarize book-level trade volume this week.
-
Which customers have unhedged positions?
-
What is my current margin requirement?
-
Show mark-to-market values by customer.
-
Who traded the most JPY options today?
-
Compare today’s exposure with yesterday’s.
-
Are all trades in the FX-G10 book within limits?
-
Show active positions set to expire this week.
-
Who has the largest risk in USD/CHF?
-
What is my unrealized gain or loss in FX-EM?
-
Which customers opened EUR/JPY trades?
-
Give me my top 5 risk contributors today.
-
Show new trades from Cust104.
-
List books where USD exposure increased.
-
Who placed the most leveraged trade today?
-
Did anyone trade during the Tokyo session?
-
What’s the average position size today?
-
Which book has the highest open interest?
-
Who added or closed positions today?
-
Compare FX-G10 and FX-EM performance.
-
What’s the total exposure to Latin American currencies?
-
Identify customers breaching internal risk policy.
-
Show my desk’s net exposure by currency.
-
Who canceled trades today?
-
Show open forward contracts by maturity.
-
List all options exercised today.
Risk Analyst Prompts (30+ Examples)
-
List customers over VAR threshold.
-
Show current exposure by trading desk.
-
Who’s carrying risk over the weekend?
-
Compare risk limits with actual exposure.
-
Who received a margin call today?
-
Summarize today’s risk alerts.
-
Show books with leverage > 5x.
-
Which currencies had the largest position delta?
-
Identify books impacted by USD volatility.
-
Show all positions tied to external events.
-
Which customers have highly correlated positions?
-
List trades triggering compliance warnings.
-
Who added high-risk exotic currency positions?
-
Show position decay trends over the last 7 days.
-
Which books would be affected by a Fed rate change?
-
What are the most volatile books right now?
-
Highlight trades not fully collateralized.
-
Identify risk clusters by customer type.
-
Show the delta of options by currency.
-
Forecast exposure in case of 100bps USD move.
-
Who is operating near their credit limit?
-
Are any VAR models out of tolerance?
-
Show drawdown trends by region.
-
Highlight suspicious trade patterns.
-
Show risk-weighted exposure by sector.
-
What would happen if EUR drops 2%?
-
Identify gaps in risk coverage.
-
Show customer concentration risk.
-
What’s the exposure to Chinese Yuan?
-
List traders with over-limit activity.
Onboarding and Compliance Ops Prompts (30+ Examples)
-
Who is pending onboarding?
-
Show onboarding timeline for Cust302.
-
Who passed AML but failed credit check?
-
List customers cleared for spot but not options.
-
Who has missing tax documentation?
-
Show onboarding SLAs per region.
-
List all KYC documents still pending.
-
Who was manually approved?
-
Which accounts are in review?
-
Show the average onboarding time.
-
List high-risk customers approved this week.
-
Identify duplicates in onboarding logs.
-
Who missed the onboarding deadline?
-
Which customers are waiting on legal review?
-
Show latest onboarding comments.
-
Who is cleared for forwards only?
-
Who failed sanction screening?
-
Show onboarding health by book.
-
List customers needing updated docs.
-
Show margin agreement status by customer.
-
Who is blocked from trading today?
-
Show onboarding dropouts.
-
List all customers in FX-G10 book with incomplete KYC.
-
Who had risk override approved manually?
-
Which customers were escalated to compliance?
-
Who is restricted to specific currency pairs?
-
Show approval workflow logs for Cust1005.
-
Identify customers without ID verification.
-
Show time to trade approval per customer.
-
List onboarding cases reopened after closure.
Tech Stack
| Layer | Tools |
|---|---|
| Auth & Roles | AWS Cognito |
| API Gateway | Lambda + FastAPI |
| LLM Inference | VLLM (on G5 spot) and AWS Bedrock |
| Models | Claude, Titan, LLaMA, Mixtral, T5 |
| Storage | S3 + Parquet + Athena |
| Mapping | DynamoDB |
| Caching | Redis (ElastiCache) |
| Monitoring | CloudWatch + Prometheus |
Prompt Access Control
Each user’s JWT token includes:
The agent injects context automatically:
Cost Optimization Strategies
| Component | Optimization |
|---|---|
| GPU Hosting | G5 Spot with autoscaling |
| Model Routing | Use Haiku/Titan for standard Q&A, Sonnet for heavy prompts |
| RAG Indexing | Nightly refresh only |
| Storage | S3 (Parquet) with Athena instead of Redshift |
| Prompt Caching | Redis for high-frequency prompts |