Skip to content
Nguyen Le edited this page Nov 16, 2026 · 8 revisions

PortfolioGuard AI

Team 62 · AlphaForge AI Datathon 2026

Your AI agent can trade. PortfolioGuard makes it show its work first.

Before you buy or sell, five cooperating AI agents check whether you are reacting to the market or to your own fear, and show you the evidence. Every figure comes from deterministic code in AWS Lambda; language models only route, read headlines and explain.

Simulation only. Fictional portfolios. Not financial advice.


How it works

Step What happens Where
1 · Connect Your own AI agent (Claude, Muse, Dots) sends a low, mid or high risk persona through an MCP connector MCP server → Intake agent
2 · Adapt The account type adjusts it: an RRSP lowers the risk one step Intake agent
3 · Compare Two market snapshots, before and after the refresh, show exactly what moved Market agent → Lambda over S3
4 · Classify The move is scored company-driven, market-driven or hype, relative to that stock's normal swing, your profile and how big the position is in your portfolio Classifier agent → Lambda
5 · Warn A pop-up gives the short verdict before you act; the chat holds the full evidence Explainer agent → Bedrock

Explore

Page What you will find
Problem Why investors act on fear, and what is missing today
Solution The user journey and what makes it different
Architecture System layers, the request path, deployment
Agents The five AgentCore agents, routing, preloading
Trade Check Signals, scores, thresholds and persona-aware verdicts
Technologies Stack, MCP connector, guardrails, validation, limits
Impact Guarantees and outcomes
Case · Team The brief and the people

Status

Component State
Five agents on Amazon Bedrock AgentCore (Orchestrator, Intake, Market, Classifier, Explainer) Deployed, us-east-1
Deterministic calculations (AWS Lambda over versioned S3 data) Deployed
Persona-aware, volatility-aware, position-size-aware verdicts (trade-guard-v4) Built and deployed (Lambda); agent package rebuild pending
Dashboard: Holdings pop-up, chat, execution trace with AWS runtime names Built
Connect Claude (remote MCP server) Built; see Technologies
AgentCore observability dashboard (traces, token cost) Not enabled yet

Clone this wiki locally