🛡️ Fiduciary Agent: The Risk-Aware Orchestrator

1. The Pitch

The Problem: Most AI agents are "Yes-Men", they prioritise politeness over business rules. In e-commerce, an agent that approves every refund request can bankrupt a company. Simple chatbots lack the context to distinguish between a VIP customer and a fraudulent claim.

The Solution: The Fiduciary Agent. This is not a chatbot; it is an Authorised RPA Controller. It acts with "Fiduciary Responsibility," calculating Customer Lifetime Value (CLV) and assessing risk before executing any financial transaction. It empowers Gemini-Flash-Lite to say "No" when necessary, protecting the enterprise bottom line.

2. Architecture: Brain, Hands, & Memory

The system moves beyond simple conversation to a strict "Check-then-Act" Protocol:

• 🧠 The Brain (Gemini 2.5 Flash Lite): Acts as the "Controller." It does not make up answers; it strictly follows a sequence to invoke the correct tools.
• 🦾 The Hands (Custom Tools): Python-based governance layer. The business logic (Math, Risk Ratios, Bank Balances) is hidden here. The tool physically blocks the agent from processing risky refunds.
• 💾 The Memory (Session Service): Maintains the user's identity (judge_01) and context across turns, allowing the agent to handle multiple requests in a single session without amnesia.

3. Concepts Demonstrated

This submission explicitly implements three core concepts from the 5-Day AI Agents Intensive Course with Google:

1. Agent Powered by an LLM: Leveraging gemini-2.5-flash-lite for high-speed, low-latency reasoning.
2. Custom Tools: Implementation of execute_refund which returns rich objects (Status + CLV + Risk Flag), not just text strings.
3. Sessions & State Management: Using InMemorySessionService to persist the conversation state and user identity.

4. Setup

• Framework: Google Agent Development Kit (ADK)
• Model: gemini-2.5-flash-lite (Optimised for Edge/Speed)
• Logic: Enterprise Risk Governance (CLV > Refund Value)