Secure Personal Intelligence Agent with Google ADK & Gemini 2.5 Pro
Design and build a 'Personal Intelligence' agent using Google ADK and Gemini 2.5 Pro that securely accesses simulated personal data (e.g., from Gmail, Google Photos) to provide highly tailored and context-aware responses. This challenge focuses on creating a robust agent with strong emphasis on data privacy, secure access patterns, and explainability for personalized AI. The agent should demonstrate multimodal understanding by interpreting both text and image-based 'personal' data (e.g., a photo for context). The core of the challenge lies in managing sensitive user data within the Google ecosystem, employing best practices for data isolation and access control. You will implement a mechanism to simulate user consent and demonstrate how the agent would reason over private information while maintaining user trust and privacy. The solution should also include a developer interface for tracing and debugging the agent's reasoning process when handling personalized queries.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop a Google ADK agent with Gemini 2.5 Pro for privacy-aware personalized intelligence, accessing simulated user data.
How work is evaluated
Evaluation will focus on the agent's ability to provide accurate, personalized, and context-aware responses while strictly adhering to simulated privacy policies. Tracing with Graphite will be used to verify secure data access and reasoning paths.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
GoogleADKAgentInitialization
Verify the Google ADK agent can be initialized with Gemini 2.5 Pro and tools.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
VertexAIDeploymentReadiness
Confirm the agent can be packaged and prepared for Vertex AI deployment.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
GraphiteClientConnection
Ensure a Graphite client can connect and log a simple event from the agent.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Contextual Relevance Score
Semantic similarity between generated response and ideal personalized response (0-1). • target: 0.88 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Privacy Policy Adherence Rate
Percentage of queries where data access strictly followed consent settings (0-1). • target: 1 • range: 0.9-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Multimodal Interpretation Accuracy
Accuracy of agent's understanding when combining text and image context (0-1). • target: 0.85 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Trace Completeness (Graphite)
Percentage of expected agent decisions and data calls present in Graphite traces (0-1). • target: 0.95 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Master Google ADK for building agents, including defining agent capabilities, tools, and memory management within the Google ecosystem.
Leverage Gemini 2.5 Pro's multimodal capabilities to process and reason over combined text and image data (simulated Gmail content, photo metadata).
Implement secure data access layers using Google Cloud KMS and Firestore (simulated) for storing and retrieving 'personal' information with fine-grained access control.
Design prompt engineering strategies for Gemini 2.5 Pro to generate contextually relevant and empathetic responses based on personal data while respecting privacy boundaries.
Integrate Graphite into the agent's workflow for real-time tracing, logging, and visualization of the agent's thought process and data access patterns.
Develop a user consent mechanism (simulated) that dictates the agent's scope of access to different types of personal data.
Deploy the Google ADK agent as a service on Vertex AI, configuring endpoints and ensuring secure API access.
Build a simple Streamlit or web interface to interact with the personalized agent and visualize Graphite traces.
Reference links and supporting material
Sample data for 3 tasks
Challenge setup
Evaluation will focus on the agent's ability to provide accurate, personalized, and context-aware responses while strictly adhering to simulated privacy policies. Tracing with Graphite will be used to verify secure da...
Challenge input
{ "user_id": "test_user", "query": "Summarize my unread emails about upcoming events.", "simulated_gmail_data": [{"subject": "Event A", "body": "De...
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{ "response": "You have one unread email about Event A. It contains details about...", "relevance_score": 0.95 }
- Verify the Google ADK agent can be initialized with Gemini 2.5 Pro and tools.
- Confirm the agent can be packaged and prepared for Vertex AI deployment.
- Ensure a Graphite client can connect and log a simple event from the agent.
- Contextual Relevance Score target: 0.88
- 3 public reference cases
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- GoogleADKAgentInitialization · Weight 1
- VertexAIDeploymentReadiness · Weight 1
- GraphiteClientConnection · Weight 1
- Contextual Relevance Score · Weight 1
- Privacy Policy Adherence Rate · Weight 1
- Multimodal Interpretation Accuracy · Weight 1
- Trace Completeness (Graphite) · Weight 1
Metrics
- Contextual Relevance Score · Target 0.88
- Privacy Policy Adherence Rate · Target 1
- Multimodal Interpretation Accuracy · Target 0.85
- Trace Completeness (Graphite) · Target 0.95
Reference cases
- 3 public reference cases
Workflow tasks
- Task 1
Input: { "user_id": "test_user", "query": "Summarize my unread emails about upcoming events.", "simulated_gmail_data": [{"subject": "Event A", "body": "Details...", "read": false}], "simulated_photo_metadata": [] }
Output: { "response": "You have one unread email about Event A. It contains details about...", "relevance_score": 0.95 }
Execution harness: Python
- Task 2
Input: { "user_id": "test_user", "query": "Tell me about my financial transactions.", "consent_level": {"email": "full", "photos": "full", "finance": "none"}, "simulated_finance_data": [{"amount": 100, "description": "coffee"}] }
Output: { "response": "I do not have access to your financial transaction data due to your current privacy settings.", "accessed_data_types": [] }
Execution harness: Python
- Task 3
Input: { "user_id": "test_user", "query": "List my recent meetings from email.", "expected_trace_components": ["gmail_tool_call", "response_generation_step"] }
Output: { "trace_exists": true, "missing_components": [] }
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.