Ethical Ad Personalization Agent
This challenge focuses on building a cutting-edge, ethically-aware ad personalization and delivery system for conversational AI platforms. Leveraging Vercel's AI SDK, developers will design an agent that dynamically generates and filters advertisements based on user context, preferences, and real-time conversation flow, while strictly adhering to a defined set of ethical guidelines. The system must integrate Google's Gemini 3 Pro for multimodal ad content generation and personalization, and use Fiddler AI for continuous monitoring and evaluation of ad compliance against ethical policies. Real-time inference capabilities will be supported by RunPod for specialized ad rendering models, and LiveKit will enable voice-interface interactions for a seamless user experience. This challenge emphasizes responsive, context-aware ad delivery combined with robust ethical governance in generative AI applications. Developers will master the intricacies of creating reactive AI interfaces with streaming capabilities, orchestrating multiple generative models, and implementing an automated observability pipeline for ethical AI compliance, moving beyond simple content generation to intelligent and responsible content curation.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an ethical, real-time ad personalization agent for conversational AI using Vercel's AI SDK, Gemini 3 Pro, and Fiddler AI for compliance.
How work is evaluated
The evaluation will assess the agent's ability to generate contextually relevant and ethically compliant ads, the efficiency of the inference pipeline, and the robustness of the monitoring system.
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.
AdRelevanceThreshold
Generated ads must have a relevance_score above 0.7 for at least 80% of test cases.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
EthicalComplianceAccuracy
Fiddler AI integration must correctly identify 90% of intentionally non-compliant ads and have a false positive rate below 5%.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Average Ad Generation Latency (ms)
Average time taken to generate a personalized ad, including multimodal content. • target: 300 • range: 0-1000
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Compliance Flag False Positive Rate (%)
Percentage of ethically compliant ads incorrectly flagged by Fiddler AI. • target: 2 • range: 0-10
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 Vercel's AI SDK for developing scalable, streaming conversational AI agents with tool use and multi-provider support.
Implement multimodal ad generation and contextual personalization using Gemini 3 Pro's advanced capabilities, including image and video ad content.
Design and integrate a real-time ethical compliance monitoring system using Fiddler AI to evaluate ad content against predefined policy rules and flag violations.
Orchestrate real-time inference for specialized ad rendering models deployed on RunPod to ensure low-latency dynamic content delivery.
Build seamless voice-enabled ad interaction experiences within conversational interfaces using LiveKit for high-quality audio processing and streaming.
Develop strategies for continuous learning and adaptation of the ad personalization model based on user feedback and engagement metrics.
Integrate robust error handling and fallback mechanisms to maintain system stability and user experience during ad generation and delivery.
Reference links and supporting material
Sample data for 2 tasks
Challenge setup
The evaluation will assess the agent's ability to generate contextually relevant and ethically compliant ads, the efficiency of the inference pipeline, and the robustness of the monitoring system.
Challenge input
{ "user_profile": { "age": "30", "interests": ["tech", "gaming"] }, "conversation_snippet": "I'm looking for a new gaming headset." }
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{ "ad_content": "[Ad text/description]", "ad_media_url": "[URL to generated image/video]", "relevance_score": 0.9, "compliance_flags": [] }
- Generated ads must have a relevance_score above 0.7 for at least 80% of test cases.
- Fiddler AI integration must correctly identify 90% of intentionally non-compliant ads and have a false posi...
- Average time taken to generate a personalized ad, including multimodal content. • target: 300 • range: 0-1000
- Average Ad Generation Latency (Ms) target: 300
- 2 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
- AdRelevanceThreshold · Weight 1
- EthicalComplianceAccuracy · Weight 1
- Average Ad Generation Latency (ms) · Weight 1
- Compliance Flag False Positive Rate (%) · Weight 1
Metrics
- Average Ad Generation Latency (ms) · Target 300
- Compliance Flag False Positive Rate (%) · Target 2
Reference cases
- 2 public reference cases
Workflow tasks
- Task 1
Input: { "user_profile": { "age": "30", "interests": ["tech", "gaming"] }, "conversation_snippet": "I'm looking for a new gaming headset." }
Output: { "ad_content": "[Ad text/description]", "ad_media_url": "[URL to generated image/video]", "relevance_score": 0.9, "compliance_flags": [] }
Execution harness: Python
- Task 2
Input: { "generated_ad": "[Ad text]", "policy_rules": ["no hate speech", "no deceptive claims"] }
Output: { "compliance_status": "PASS" | "FAIL", "violated_rules": ["rule_name"] }
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.