Project overview
The challenge
- Marketing teams needed to select highly specific audiences for 50+ events/year but relied on static, outdated lists and manual SQL queries from IT.
- Engagement data (webinars, downloads) was siloed from profile data, making behavioral targeting impossible.
- Manual content creation for each event (LinkedIn posts, emails) was a bottleneck, reducing the number of events that could be promoted.
- No feedback loop existed to refine audience selection based on past performance or 'lookalike' criteria.
Marketing teams at B2B SaaS companies spend 3–5 days per campaign crafting audience segments — cross-referencing attendee history, firmographics, and engagement scores before a single email sends. This n8n orchestration collapses that cycle to under 3 minutes: a marketer describes the audience in plain language, and a team of specialised AI agents (Orchestrator, Keyword Generator, SQL Validator) translates the intent into validated SQL CTEs against a live PostgreSQL database — no analyst, no ticket, no waiting.
The targeting logic encodes a 1 000-point scoring model combining behavioural signals (event attendance, content interaction, session replays) with profile fit — generating audience cohorts of previously impossible granularity. A Perplexity-backed keyword diversification agent produces 150+ query variants to maximise recall without sacrificing precision.
Approval and output flow through Microsoft Teams Adaptive Cards, keeping stakeholders in their existing tooling. Once approved, the system auto-generates LinkedIn post copy and visual banners (Gemini + OpenAI), then dispatches the campaign via Mailjet — turning a multi-day analyst-and-designer loop into a single supervised run.
Technical architecture
n8n Orchestrator: Central hub managing the state and flow between agents, database, and user interfaces.
AI Agent Swarm: A network of specialized LLM agents (using LangChain nodes in n8n) handling specific tasks: understanding intent, generating keywords, writing SQL, and validating safety.
SQL Proxy Layer: A custom secure middleware ensuring AI-generated SQL is read-only and performance-optimized before hitting the production database.
Human-in-the-Loop: Critical approval steps integrated via Microsoft Teams Adaptive Cards, allowing non-technical users to validate audience segments before campaign launch.
Key challenges & solutions
Natural Language to Reliable SQL
Solved the hallucination problem by implementing a 3-step verification process: Keyword Expansion → SQL Generation → Syntax Validation Agent. The system safely handles complex logic like 'people who attended X but not Y'.
Secure AI-DB Interaction
Built a custom SQL proxy middleware to sanitize inputs and enforce read-only access, preventing any risk of SQL injection or data corruption from AI-generated queries.
Handling Abstract Scoring Logic
Translated a subjective 1,000-point business scoring model into concrete, performant SQL Common Table Expressions (CTEs) that calculate scores in real-time across millions of rows.
Complex Multi-Step Approvals
Engineered a robust state management system within n8n to handle long-running workflows that pause for human approval on Teams (waiting for webhook callbacks) without timing out.
Impact & results
3 days of marketing analyst work — audience research, SQL segmentation, copy drafting, banner briefing — compressed to under 3 minutes per campaign
Zero SQL knowledge required: non-technical marketers access complex PostgreSQL audience data through plain-language natural-language queries
1 000-point behavioral scoring model enables audience cohorts previously impossible with static lists or manual segmentation
End-to-end content supply chain automated: audience selection → LinkedIn copy → visual banners → email dispatch, all in a single supervised run
Key features
- Multi-Agent AI Orchestration (GPT-4o, Gemini, Perplexity)
- Natural Language to SQL conversion with validation
- 1,000-point audience scoring algorithm
- Interactive Teams Adaptive Cards for approval
- Automated LinkedIn post & banner generation
- Secure SQL Proxy for database interaction
- Dynamic audience expansion logic (threshold reduction)
- Keyword diversification agent (150+ variants)
- Automated email marketing orchestration (Mailjet)
- Complex n8n workflow with subgraph architecture
