🚀 Privacy-First Personal Assistant

• Track: Concierge Agents.

• Subtitle: A Capstone Project for the Kaggle 5-Day AI Agents Intensive Course. This agent protects user privacy by asking for explicit permission before saving any long-term memories.

---

📺 Project Video

Youtube Video

1. 🎯 The Pitch (Problem, Solution, Value)

This project maps directly to the "Category 1" evaluation criteria.

• The Problem: Standard personal assistants have a major flaw: they save everything you say by default. This creates a privacy nightmare, as users have no control over whether sensitive information (like health data, family birthdays, or private thoughts) is stored permanently.

• The Solution: This agent inverts the model. It's a "Privacy-First" assistant that assumes all information is temporary. It is built to detect sensitive information and must ask for explicit user permission before saving anything to its long-term memory.

• The Value (Core Concept): This project's core idea is building user trust through explicit consent. It gives the user genuine control over their data by using a Human-in-the-Loop (HITL) workflow. The agent can still be a helpful, stateful assistant, but only with facts the user has personally approved.

2. 🛠️ The Implementation (Architecture & Key Concepts)

This agent's architecture and code are designed to satisfy the "Category 2: The Implementation" criteria by using 3+ key concepts from the course.

🏛️ Architecture

The agent's workflow is centered around a "pause and resume" logic. The full process flow and system architecture are visualized below:

🔑 Key Concepts Used

This project implements three core concepts from the course notebooks:

1. Long-Running Operations (LRO): This is the core of the privacy feature. The agent uses a custom tool, process_user_fact, which detects sensitive keywords (e.g., "allergic," "birthday"). When a sensitive fact is found, the tool calls tool_context.request_confirmation() to pause the entire agent workflow and wait for the user's "yes/no" response. This is the "Human-in-the-Loop" (HITL) pattern.

2. Long-Term Memory: This concept is split into two parts:

   • Saving: If (and only if) the user approves the LRO, the agent resumes, and the tool saves the conversation to the InMemoryMemoryService using await memory_service.add_session_to_memory().
   • Retrieving: The agent uses the preload_memory tool. This proactively loads all approved facts from long-term memory into the agent's context at the start of every new conversation.

3. Context Engineering (Compaction): This is the "forget" mechanism. If the user rejects a save, the fact is not saved to long-term memory. We also enable EventsCompactionConfig on the App with an aggressive interval. This ensures the rejected fact is quickly and automatically removed from the short-term session history, making the agent truly "forget" it.

✨ Bonus Concepts Used

• Effective Use of Gemini: The agent's reasoning is powered by Gemini(model="gemini-2.5-flash"), satisfying the bonus criteria.
• Custom Tools: The entire project is orchestrated by a custom FunctionTool (process_user_fact) that handles the complex LRO and memory logic.

3. 💻 Technology Stack

• Google Agent Development Kit (ADK)
• Google Gemini API (gemini-2.5-flash)
• Python 3 & asyncio
• Google Colab (for development and demonstration)
• Deployment: Agent Engine (ADK) on Google Cloud (Vertex AI)

4. ⚙️ Setup & Usage

This project is designed to run in a Google Colab or Kaggle Notebook.

Setup

1. Get API Key: Create a Google Gemini API key at Google AI Studio.
2. Add Secret: Open the notebook in Google Colab and click the "Secrets" (🔑) tab. Create a new secret named GOOGLE_API_KEY and paste your key as the value.
3. Run Setup Cells: Run cells 1-6 in the notebook. This will:
   • Install google-adk.
   • Import all libraries.
   • Configure the API key.
   • Define the LRO tool, the agent, the app, and the helper functions.

How to Run the Interactive Demo

Run the final cell (Cell 8: FINAL DEMO: Interactive Chat Loop) to start the agent.

Note: The demo code (cell 7) is a pre-scripted run to show all features working. The interactive chat (cell 8) is for live demonstration.

You can test the full privacy workflow:

1. Test 1: Save a Fact

   • You: I am allergic to strawberries.
   • Agent: (PAUSES) ... Assistant requires approval. Approve? (y/n):
   • You: y
   • Agent: Got it. I've saved this to your long-term profile.

2. Test 2: Reject a Fact

   • You: My birthday is August 10th.
   • Agent: (PAUSES) ... Assistant requires approval. Approve? (y/n):
   • You: n
   • Agent: Okay, I will not save this fact.

3. Test 3: Verify Memory

   • Restart the interactive chat (cell 8) to simulate a new session.
   • You: What do you know about me?
   • Agent: I know that you are allergic to strawberries.
   • (Notice: It correctly remembers the allergy but not the birthday).

---

This project was built for the Kaggle 5-Day AI Agents Intensive Course (Nov 2025).