▸System Overview
phase 1 · human-in-the-loopKairon is an AI-supervised trading platform. A TradingView strategy fires a webhook; an AWS-hosted AI agent validates the signal against five deterministic risk filters; the proposal lands in this terminal for the owner to approve; on approval, a Lambda places the order on OANDA. A reconciler reads broker fills back into the journal every five minutes. The system is currently wired to the OANDA practice account — no real capital is at risk.
▸Data Flow
signal → validation → approval → execution → reconciliation▸The Validator Agent
strands + bedrock opus 4.7Every webhook is forwarded to a Bedrock AgentCore runtime running a Strands agent on Claude Opus 4.7. The agent has access to five deterministic tools — not free-form decision making. It must call them, read their outputs, and emit a verdict of APPROVE or BLOCK with reasoning.
Median validation latency: ~13s. Tool calls per signal: 4–5. Cost: ~$0.05–$0.15 per validation.
▸Why Human-in-the-Loop
phase 1 controlThe agent recommends, the owner decides. APPROVE proposals are visible in this terminal; the executor only places an OANDA order after a click. BLOCKED signals are persisted with reasoning for audit but never reach the order book.
Phase 2 (full autonomy) is gated on 30+ closed paper trades, Sharpe ≥ 1.0, max drawdown < 15%, and a clean run through a high-volatility news event.
▸Risk Filters
every signal must clear all five▸Build Status
updated 2026-06-11- [✓]Phase 1 — AWS infra (4 CDK stacks)
- [✓]Phase 2 — OANDA API wrapper
- [✓]Phase 3 — Reconciliation loop + webhook receiver
- [✓]Phase 4 — Approval UI on CloudFront
- [✓]Phase 5 — Validator agent (Strands + Opus 4.7)
- [~]Phase 6 — 1-month paper trading on OANDA practicein progress
- [ ]Pre-go-live — Cognito auth, alarms + canaries, cost anomaly detectionqueued
- [ ]Phase 2 future — Interactive Brokers integration, full autonomyqueued