Marketing Agency Platform

The problem

Brand knowledge, past campaigns and customer feedback often sit in different documents and conversations. When that context gets left out of a brief, even well-written content can miss what the business actually needs.

The solution

I built Headcount as an end-to-end agency platform, from analyzing a client’s business and customers through strategy, campaign planning, content creation, review and scheduling. The client’s profile, brand guidance and feedback stay connected to the work throughout.

An orchestrator turns requests into plans, assigns work to specialist AI agents and saves what they produce. People review the strategy, facts and brand fit before content is approved and scheduled.

How it works

The architecture, at a glance.

  1. Client & customer analysis

    Business context, customer intelligence and brand guidance

  2. Strategy & planning

    An orchestrator builds a plan and assigns specialist agents

  3. Content creation

    Specialist agents produce articles, images and social content

  4. Review & scheduling

    Check strategy, facts and brand fit, then approve and schedule

Client context and previous content stay connected to each request. Review feedback is saved for future improvements.

Individuals and teams remain responsible for strategy, approval and launch decisions.

Technology

Application
Next.js, React and TypeScript for the client workspace and operational interface.
Data & contracts
PostgreSQL, Prisma, tRPC and schema validation for jobs, agent runs and execution plans.
Marketing planning & specialist agents
Grok 4.3 handles request analysis and execution planning, with specialist agents for brand strategy, customer personas, content strategy, articles, SEO and social posts. Each agent receives the relevant client context and produces structured output that can be checked and saved.
Composer
Claude Sonnet 4.6 is the configured default for the conversational Composer workflow and summaries of completed work. It connects the conversation to the platform’s marketing workflows.
Image generation
GPT Image 2 generates images, with GPT Image 1.5 configured as a fallback. Model choices can be overridden through configuration.

Worth mentioning

Giving agents the right context

I connected each request to the client’s profile, deep customer intelligence, marketing context and feedback. The agents work from the same business information as they plan and produce the work.

Checking the plan before running it

The orchestrator’s plan is checked against a defined schema. I save each job’s progress and outputs so it’s possible to see what happened when a model response fails or a request needs another pass.

Making room for judgment

I built review into the workflow so people can check whether the work fits the strategy, supports its claims and sounds like the brand. Their feedback is saved so it can inform future improvements to the workflow.