Apex — Trade & Risk Intelligence Platform

An AI-powered incident intelligence platform for trading operations. Eliminates 20–45 minutes of manual context gathering per incident by pre-assembling a complete operational briefing the moment a trade break or risk limit breach alert fires.

---

The Problem

When a trade break or risk limit breach alert fires, operations teams manually reconstruct context across 4–6 systems — trade records, position blotters, risk limit dashboards, counterparty records, regulatory obligation references — before they can act. This takes 20–45 minutes per incident under normal conditions. Under pressure, with multiple alerts firing simultaneously, things get missed.

Apex eliminates the manual gathering step. The alert arrives → Apex assembles the full context automatically → the operations team receives a ready-to-act briefing in seconds.

---

Architecture

External alert source
(Grafana / PagerDuty)
        │
        ▼ POST /analyse
  Internet-facing ALB
  (IP-restricted listener rule)
        │
        ▼
  FastAPI on ECS Fargate
        │
        ├─► SQS Queue ──► Lambda ──► SES notification
        │
        ▼
  AI Pipeline
  ┌─────────────────────────────────────┐
  │                                     │
  │  1. RAG pipeline step               │
  │     Cohere Embed v4 + Rerank 3.5    │
  │     S3 Vectors                      │
  │                                     │
  │  2. Reasoning agent                 │
  │     Claude Sonnet 4.6               │
  │     5 tools (structured data)       │
  │                                     │
  │  3. LLM-as-judge (async)            │
  │     Claude Haiku 4.5                │
  │     4 scoring dimensions            │
  │                                     │
  └─────────────────────────────────────┘
        │
        ▼
  DynamoDB (incident store)
        │
        ▼
  React + Vite Frontend
  (S3 + CloudFront)

Infrastructure: All AWS. Terraform with S3 backend (native locking). GitHub Actions with OIDC authentication (no long-lived keys). ECS Fargate, DynamoDB, S3 Vectors, SQS, Lambda, SES, CloudFront.

---

The 8 Demo Scenarios

#	Alert type	Instrument	Desk	Complexity
S01	Trade break — price mismatch	TSLA	Equity Derivatives	Low
S02	Trade break — settlement failure	UK Gilt 4.25% 2034	Fixed Income	Low
S03	Position limit breach — float %	NVDA	Equity Derivatives	Medium
S04	VaR breach — tech basket	NVDA/AAPL/MSFT/GOOGL/META	Equity Derivatives	Medium
S05	Counterparty credit breach	Goldman Sachs	Equity Derivatives	Medium
S06	Margin call	E-mini S&P 500 Dec 2025	Macro Rates	High
S07	Cascading trade break	AAPL	Equity Derivatives	High
S08	Sector concentration + UCITS regulatory flag	XOM/CVX/SLB	Energy & Commodities	High

---

Key Design Decisions

RAG as a pipeline step, not an agent tool. Regulatory document retrieval happens before the agent runs and is injected into the system prompt. This keeps latency low and retrieval deterministic per alert type. See docs/adr/ADR-001-002-design-decisions.md.

No agent frameworks. The agentic loop is implemented directly against the Anthropic SDK. No LangChain, no LlamaIndex. This keeps the pipeline transparent and the dependencies minimal.

5 tools, no more. The reasoning agent has exactly 5 structured data tools. Tool count is a design decision — more tools means longer loops and more opportunity for drift.

Mock data only. All trading data is synthetic. There are no real trading system connections. The fixture data is realistic enough to demonstrate genuine domain knowledge.

---

Domain Context

Trade breaks occur when there is a mismatch between what Apex's firm recorded for a trade and what the counterparty (e.g. Goldman Sachs, JPMorgan) confirmed. Unresolved breaks can become failed settlements, triggering regulatory reporting obligations and counterparty relationship damage.

Risk limit breaches occur when a desk exceeds its authorised exposure — whether by position size (float %), daily VaR, counterparty credit, or sector concentration. Breaches must be reported to risk management immediately.

Margin calls are urgent collateral demands from clearing houses (e.g. CME) when a leveraged position moves against the firm. If unmet within the deadline, the position is liquidated — typically at the worst possible time.

The operations team currently gathers context from: the trading system (Murex/Calypso), counterparty confirmation systems (SWIFT), the position blotter, the risk system (Bloomberg AIM/RiskMetrics), the counterparty record, and regulatory reference documents. Apex replaces those 6 manual lookups with one structured briefing.

---

Getting Started

Local development

# Backend
cd backend
pip install -r requirements.txt
APEX_ENV=local uvicorn app.main:app --reload --port 8000

# Frontend
cd frontend
npm install
VITE_API_BASE_URL=http://localhost:8000 npm run dev

# Fire a demo alert (S06 — margin call)
curl -X POST http://localhost:8000/analyse \
  -H "Content-Type: application/json" \
  -d @data/scenarios/s06_sp500_margin_call/alert.json

Infrastructure

cd infra
terraform init
terraform plan -var-file=dev.tfvars
# terraform apply — only via CI/CD pipeline

Embed the knowledge base

cd knowledge_base
python embed.py --raw-dir raw/ --chunks-dir chunks/ \
  --s3-bucket apex-vectors-$(aws sts get-caller-identity --query Account --output text) \
  --region us-east-1

---

Environment Variables

Variable	Description	Required
APEX_ENV	local / mock / production	Yes
APEX_AWS_REGION	AWS region	Yes
APEX_DYNAMODB_TABLE	DynamoDB table name	Yes
APEX_S3_VECTORS_BUCKET	S3 Vectors bucket name	Yes
AWS_ACCESS_KEY_ID	AWS credentials (local only)	Local
AWS_SECRET_ACCESS_KEY	AWS credentials (local only)	Local
VITE_API_BASE_URL	Backend URL for frontend	Build time
VITE_DEMO_MODE	Enable scenario trigger panel	Build time