SERP Surg3on: Multi-Agent SEO Optimizer

Enter your page
See the report
Agent architecture

🌟 Inspiration

Manual SEO audits are slow, error-prone, and rarely leverage the rich reasoning now possible with Google’s Agent Development Kit. We set out to prove that a swarm of small, purpose-built agents can out-analyze a human in the time it takes to reload a page.

🛠️ How We Built It

Layer	Tech	Role
Agents & Orchestration	Google ADK (Python)	Hierarchical controller → four specialist agents (SERP Collector, Page Parser, Gap Analyzer, Recommender)
Models & Tools	Vertex AI Gemini Flash & Vertex Search Extensions	RAG over competitor pages, real-time SERP, and on-the-fly Python scoring
Execution	Cloud Run (backend)	Parallelizes agent calls; retries on rate-limits
Frontend	Firebase Hosting + Flutter	Live report

🔍 What We Learned

AgentOps first: versioned prompts, automatic trajectory tests, and LLM-as-Judge scoring gave us CI/CD-like safety nets.
Design patterns matter: the diamond pattern (analysis → tone-polish → user) kept responses concise while preserving depth.
Agentic RAG beats static RAG: iterative searches lowered hallucination rates by 28 %.

🚧 Challenges

Agent coordination – in some cases prompting is extremly important to get right.
Tool selection – building custom tools vs using readily available ones.
Latency under load – Enabling tools to run in parallel, using faster, ligher models.

🤝 What’s Next

Add a Contractor Agent to auto-open issues in GitHub when critical SEO errors are found.
Publish our Agent Ops playbook as a blog post for bonus hackathon points.