Case Study · 2025
SignalDeck
An AI-assisted trading platform that coordinates twelve Claude-based strategist agents to surface trade ideas, manage risk, and execute orders — all driven by a single-table DynamoDB design and serverless AWS infrastructure.
The problem
Retail traders don't lack for tools — they lack for a coordinated system. Scanners surface hundreds of candidates, risk models live in spreadsheets, and execution happens in a third place entirely. SignalDeck compresses that whole stack into one system, where agents specialize in narrow decisions and an orchestrator composes their output into executable trades.
Architecture
A 12-agent Claude pipeline runs every 15 minutes during market hours. Each agent has a narrow job — scanning, sentiment, regime classification, risk vetting, timing — and emits a typed dataclass. The orchestrator merges results and hands off to a circuit breaker plus order validator before any execution happens.
- Frontend: React 18 + TypeScript on AWS Amplify
- API: API Gateway → Python Lambda (monolithic handler, 70+ routes)
- Data: Single-table DynamoDB with
USER#partition keys - Agents: Python on Lambda, triggered by EventBridge cron
- Execution bots: Long-running Python processes on EC2 with systemd
What I'm proud of
- Treating AI agents like disciplined coworkers: each one has a narrow job, a typed output, and gets second-opinioned before anything touches real money.
- Building a circuit-breaker-first architecture. Safety limits are enforced in code, not prompts.
- Shipping real orders against a real broker with a real P&L. The system is accountable.
Tech highlights
- 12-agent orchestrator — scanning, risk, sentiment, regime, timing, execution
- Order validator with 10 pre-execution safety checks (position sizing, PDT, duplicates, volume sanity)
- Circuit breakers on daily loss, drawdown, consecutive losses
- Audit logging of every decision with full reasoning chain
- Demo mode for unauthenticated exploration