Applied AI and Agentic Systems in Seliee

The AI layer of Seliee, the marketing platform I built at Neeblo: a team of AI agents, an MCP server for external agents, and AI features across several model providers.

Problem and scope

Marketing teams repeat the same multi-step work across many tools: finding trends, writing and publishing content, running campaigns, answering customers and reporting. AI Office automates these workflows: you give a goal, and agents plan and carry out the steps with the panel's own tools. An agent architecture was chosen because the work spans many specialist areas and needs tools, shared state and a human decision at the points that cost money or reach the public.

Architecture

• Agent roles: a manager that plans, briefs the team and summarises results without calling tools, and 19 specialists including SEO, content, ads, social, CRM, inbox, sales, data, analyst, research and developer
• Tools: the panel's own functions; the same capabilities are exposed to external agents through an MCP server with more than 230 tools
• Model routing: each agent uses its own AI account if that provider supports tool calling; otherwise the first account whose spending limit still allows the call; otherwise a self-hosted Llama, and the fallback is logged
• State: a run is a plan of steps with statuses; each task's output is stored and passed to the steps that depend on it, and steps whose inputs failed are skipped
• Integrations: OpenAI, Claude, Gemini, Perplexity and llama.cpp, plus the platform's social, ads, email and CRM modules

Security and control

• Permissions: read and write calls run as the run's owner; publish, delete and spend calls, and any unrecognised tool, are held until a person approves
• Human approval: held calls are collected into one question at the end of the step; costly media renders wait for a cost approval
• Tenant isolation: every lookup is scoped to the merchant, and IDs passed between steps are checked to belong to that merchant
• Audit records: every AI call, tool call, question and run result is logged with its status and error
• Failure scenarios: no usable AI account stops the task with a clear error, a tool call budget of 12 per task prevents loops, and automation and developer agents only propose changes that a person switches on

Evaluation

Agents are judged by task success rate (from task statuses), the rate of wrong or rejected actions (from approvals), latency, and token cost. Every metered AI call is priced per model and recorded per account per day and month, and spending limits stop calls before a budget is exceeded. Plans are validated before they run, so a publish tool cannot slip into a step without an approval.

Evidence

What can be shared publicly: the live product at seliee.si, and the AI assistant on this website. A walkthrough demo of AI Office and the MCP server, an architecture diagram, selected code samples and measured results are available on request, as far as they can be shared.