Multi-Agent Financial Decision System

Corporate finance decisions like build vs. buy, capacity expansion, or investment sizing usually need multiple lenses. Strategic fit, risk tolerance, financial returns, modeling rigor. A single LLM collapses all of that into one voice and quietly loses the adversarial check that humans bring.

• Architected a LangGraph state machine with four specialized agents: Strategy, Risk, Finance, and Modeling.
• Added an assumption-governance layer so each agent's inputs are explicit and traceable.
• Implemented adversarial Risk review. Every recommendation gets stress-tested before it's surfaced.
• Generated three-scenario outputs (Downside / Base / Upside) plus sensitivity analyses on the key drivers.
• Built evaluation metrics for decision quality: consistency, traceability, and adversarial coverage.

• Adversarial agent review caught reasoning errors in roughly 30% of test cases that a single-pass LLM missed.
• Forcing each agent to declare its assumptions made the outputs auditable, which single-prompt approaches don't give you.
• Three-scenario output shifted the conversation from 'what's the answer' to 'where's the risk', which is closer to how a real committee thinks.

Produces go / no-go recommendations with the structure of a real finance committee memo: explicit assumptions, scenario sensitivity, risk flags. A research-grade demo of how multi-agent systems can encode consulting-style structured thinking.