Inspiration
Financial disasters happen when everyone believes the same story. We asked: What if AI could argue with itself to find the truth?
Inspired by how investment firms use adversarial analysis, we built a system where 6 AI agents debate each other before reaching conclusions.
What it does
AlphaForge answers questions like "Why did NVIDIA drop and what does it mean for my portfolio?" using a 6-agent debate system:
- Research - Gathers evidence (news, prices, knowledge base)
- Event Hunter - Identifies market events
- Portfolio Storyteller - Explains impact on YOUR port
- Conspiracy Theorist - Proposes wild alternative theories (stress test)
- Devil's Advocate - Challenges every claim with objections
- Governance - Validates evidence, rejects unsupported
The Debate: Research finds evidence → 3 agents analyze in parallel → Devil's Advocate challenges them → Agents defend with evidence or concede → Governance validates → Final answer
Admin Debug View: Click 🐛 to see complete reasoning, devery agent.
Output example: "Verdict: partially_supported · 2 debat
How we built it
Backend: Python + LangGraph (agent orchestration) + AWS) + Flask API
Frontend: React + TypeScript with debug view for complete transparency
Risk Engine: GARCH volatility models + Monte Carlo simulation
Key Design:
- Evidence created by CODE (E1, E2...), agents only cite IDs
- Every number validated against cited sources
- Governance can only harden verdicts, never soften
- Max 2 debate rounds (prevents infinite loops)
Challenges we ran into
- JSON Truncation - Agents hit token limits mid-output. Fixed: bumped limits to 6000+ tokens.
- Conspiracy Agent Generated Nothing - 0 theories, 5% ompt to DEMAND theories, raised confidence floor.
- Debug View Data Transfer - 40KB+ JSON killed URL parameters. Fixed: switched to sessionStorage.
Short-term:
- Smarter evidence scoring
- Advanced governance (compliance rules, fact-checking)
- Better risk models (copulas, regime-switching)
Vision:
- Multi-asset class (bonds, crypto, real estate)
- Continuous learning from validated analyses
- Enterprise features (SSO, audit trails)
- Educational mode (jargon-free explanations)
Stack: Python · React · AWS Bedrock · LangGraph · GARCH
Built With
- amazon-web-services
- bedrock
- cloudwatch
- flask
- langgraph
- mlflow
- python
- react
- s3