One site, one local goal
Hemet Towing is a towing company in Hemet, California. The brief was direct: build one site focused on local search for the area. The site launched on July 16, 2026 with 33 unique static pages.
The visible result is a website. The more useful subject of this case study is the system behind it: an AI orchestrator directing specialized subagents through research, SEO research, planning, development, quality assurance, review, copywriting, and image generation.
I did not use one model for every task, and I did not ask one agent to produce a complete site in a single pass. I treated the work as a coordinated production system with explicit responsibilities, review gates, and a human decision-maker throughout.
My role
I led the project solo through Beetle & Frog Design. My work covered strategy, brand adherence, system design, orchestration, and quality review.
The agents produced work. I defined the workflow, reviewed each stage, and remained responsible for whether that work was good enough to continue or ship.
An orchestrator with specialized subagents
The orchestrator manages the work; specialists handle bounded tasks.
One AI orchestrator directs a fleet of subagents specialized by responsibility:
- Research
- SEO research
- Planning
- Development
- Quality assurance
- Review
- Copywriting
- Image generation
Research and planning happen before development. Their output returns to the orchestrator, where I review it against my quality standards before allowing the next stage to continue. Development, QA, and review then operate as distinct responsibilities rather than one agent evaluating its own work.
A staged system gives me explicit places to inspect assumptions and correct direction.
Model routing by task
I route models by responsibility, not by habit.
Researchers, planners, and QA agents run on GPT-5.6 Terra. Development and review agents run on DeepSeek V4 Flash through Vercel AI Gateway. Specialized image agents use GPT-5.6 Terra for image generation.
The point is not to standardize on one model. It is to choose the model used at each part of the process according to the work that agent is expected to do.
Every specialist reports back to the orchestrator. The orchestrator delegates the next review or implementation step, but it does not replace my approval. I check in throughout the run to make sure the work remains on target.
Copy as a brand constraint
Every page should be written as a page, not filled as a template.
Dedicated copywriting agents work from brand documents. Their constraint is not simply to produce enough copy for the site. The copy needs to sound human and remain on brand.
No two pages are the same. Each page receives unique content rather than repeating one local-search pattern with substituted terms. That makes copy quality part of the system design: brand adherence and page-level specificity are requirements the agents must satisfy, not cleanup left for the end.
How I evaluated it
The human never leaves the loop and guides AI through the process.
I review the research and planning output at each stage before telling the orchestrator to continue. During development, I check in regularly to confirm that the system is still working toward the intended result. At the end, I review the complete output for quality rather than treating successful generation as successful delivery.
The evaluation discipline is part of the build. Research is not accepted because an agent completed it. A plan is not approved because it is detailed. Code is not finished because it runs. Copy is not ready because it is unique. Each artifact has to meet the quality standard for its role in the whole site.
The separation between development, QA, and review also matters. The agent producing an implementation is not the only agent judging it. Independent review gives the orchestrator another signal, while my review remains the final gate.
Results
The site was built and shipped. The claims I can make now are about the system and the work it produced:
- A staged build process coordinated through one orchestrator
- Specialized agents for research, planning, implementation, QA, review, copy, and images
- Model routing by task through Vercel AI Gateway where applicable
- Human approval gates before the system advanced
- Unique pages constrained by the company’s brand documents
- A final review of the complete site before delivery
Early ranking evidence now exists: as of September 26, 2026, hemettowing.com ranks on page 1 for the query “emergency towing hemet” in organic results, below the sponsored block. Rankings shift by query, location, and time, so I present this as a dated snapshot rather than a fixed position. The screenshot is on file.
That distinction matters. The ranking snapshot is early evidence that the system is producing the intended result, not a complete measure of long-term search performance. I can show how the work was organized, how quality was controlled, what was delivered, and the position captured on September 26, 2026. The full picture will develop over time.
Principles
The human never leaves the loop
Automation increases the amount of work a system can produce. It does not transfer responsibility for the result. I review the work at defined stages, decide whether it can continue, and remain accountable for what ships.
Route models by task, not by habit
A model is part of a role inside the system. Research, planning, QA, development, review, copy, and image generation do not have identical needs. The workflow should make those choices explicit.
Copy quality is a brand constraint, not a volume game
A large site does not need interchangeable pages. Every page should have a reason to exist, sound like the company it represents, and survive a human review. Page count is not a substitute for specificity.