.png?width=271&height=189&name=Untitled%20design%20(10).png)
Custom Policy Agent Framework for Brand Protection
Zencore partners with Zefr to develop a multi-agent AI workflow to scale their content policy development on Google Cloud.
Custom Policy Agent Framework for Brand Protection
Project Location:
United States
Industry:
AdTech / Brand Suitability
Use Case:
Agentic AI, Content Moderation, Gemini Enterprise Agent Platform, Google ADK
Website:
https://zefr.com/
About ZEFR Solutions
Zefr is a brand suitability and contextual targeting company that gives advertisers control over where their ads appear across YouTube, Meta, TikTok, Snap and the open web. Its Zefr's Atrium platform combines proprietary classification models with policy frameworks aligned to industry standards, helping advertisers avoid unsafe or off-strategy placements while reaching the right audiences. Zefr's customers include Fortune 100 brands and major holding companies.
"Zencore was able to bring agentic capabilities around policy development from plan to working prototype quickly while making use of our existing infrastructure."
Jon Morra, CAIO, ZEFR
Project Overview
Scaling Content Policy Development with Agentic AI
Zefr wanted to evaluate how agentic workflows could enhance the way content policies are created, expanded, and tested. The goal was to combine automation with expert oversight to streamline policy development, surface ambiguous or edge-case content, and continuously refine classifications through iterative feedback.
To develop a prototype of this, Zefr needed a partner who could move quickly on a framework while laying the foundation for production-grade agentic systems down the line.
Key challenges included:
-
- Agentic Architecture: Designing a multi-agent system that could coordinate distinct responsibilities (search, annotation, summarization) under a single orchestrator without losing accuracy or auditability.
- Vector Search Integration: Connecting the agent workflow to Zefr's existing vector store and embedding logic so the system could perform semantic searches across relevant content corpora.
- Evaluation Framework: Establishing a baseline and a repeatable evaluation methodology so Zefr's team could measure agent performance over time and finetune with confidence.
- Production Readiness: Building the framework with a CI/CD pipeline, persistent session management, and a clear technical handover so Zefr could extend the work after the engagement.
“With Zencore, we found a business partner who not only builds outstanding solutions but also puts a high emphasis on smooth collaboration. This led to our project being a great success - not only in terms of quality but also in terms of transparent communication and project management. The expertise in Google Cloud is remarkable and the Engineers made sure that we also learn along the journey. Working together was a pleasure and it would be great to join forces again in the future!”
Scope of Work
Zencore Delivers a Multi-Agent Foundation on Vertex AI
Zencore partnered with Zefr to design and implement a custom Policy Agent application based on Zefr's policy agent design, using Google's Agent Development Kit (ADK) and the Gemini Enterprise Agent Platform. The five-week engagement was structured around two milestones: foundational agent development with a working chat UI, followed by deployment, evaluation, and fine tuning on Google Cloud.
The application was built around four coordinated agents:
- Coordination Agent (CA): the root agent, responsible for orchestrating the workflow and handling direct user interaction.
- Search Agent (SeA): generates and executes semantic searches against Qdrant to surface content relevant to a given policy.
- Annotation Agent (AA): annotates search results against policy criteria.
- Summary Agent (SuA): synthesizes the annotated results into a clear, reviewable summary for policy teams.
Technologies
- Google Cloud Platform (GCP): Core cloud infrastructure for the deployed POC.
- Google Agent Development Kit (ADK): Framework for building the multi agent application.
- Gemini Enterprise Agent Platform: The overarching platform for building, deploying, governing, scaling, and optimizing agents.
- Agent Platform Sessions: Persistent session management for stateful, multi turn agent conversations.
- Agent Platform Evals: Integrated evaluation tooling to baseline and improve agent performance.
- CI/CD Pipeline: Automated build and deploy pipeline for the agentic application.
Business Impact
The Custom Policy Agent Framework provided Zefr a working blueprint for applying agentic AI to one of the most labor-intensive parts of its platform: developing and refining content policies. With a deployed application, an evaluation framework, and a documented architecture in hand, Zefr's team is positioned to scale the approach across additional policy domains and continue iterating on the agents with confidence.
More broadly, the engagement validated the Gemini Enterprise Agent Platform and Google's Agent Development Kit as a viable foundation for Zefr's future agentic initiatives, accelerating the path from experimentation to production for content governance use cases that demand both speed and rigor.
Key Achievements
- Multi-Agent Application Delivered: A functional chat-based application with all four agents (CA, SeA, AA, SuA) implemented and queryable both independently and as a coordinated workflow.
- Vector Database Integration: The Search Agent successfully integrated with Zefr's existing vector store and embedding logic, enabling semantic search over Zefr's content corpora.
- Deployed on GCP with Persistent Sessions: The application was deployed to a Google Cloud project with a working CI/CD pipeline and integrated with Gemini Enterprise Agent Platform Sessions for conversation persistence.
- Evaluation Baseline Established: An initial evaluation test set was built for each agent and integrated with GCP GenAI evaluation services, with two rounds of workflow and prompt fine tuning showing measurable improvement over baseline.
- Technical Design Documentation: A full technical design document was delivered and approved by Zefr's team, with clear recommendations for next-stage development.