Relay CRM — AI-Native Support System

Relay CRM is an MVP for AI native customer support: an agent that can search a knowledge base, open tickets, escalate to humans, and keep operators in the loop, with full observability on every run.

Problem

Support teams drown in repetitive questions while tribal knowledge lives in docs nobody searches. I wanted a system that could:

1. Answer from a knowledge base when confidence is high
2. Create and route tickets when it isn't
3. Detect SLA risk and alert humans before things slip
4. Stay inspectable: every tool call and model choice traced

Approach

The core is an agentic loop across multiple LLMs in Go. Different models handle different steps (planning, tool use, synthesis) behind a shared tool interface:

• Full text search over a knowledge base
• Ticket creation and human escalation
• Rules configured in YAML for automated ticketing and SLA breach detection
• Telegram alerts for operators on call
• LangSmith tracing across agent runs and tool spans

PostgreSQL is the system of record for tickets, queue state, and config. Rules stay in YAML so product can tune thresholds without a redeploy of agent logic.

What shipped

• End to end agent loop with tool use and escalation
• Live operator queue for admins
• SLA breach detection with Telegram push
• Observability that made failed runs debuggable instead of opaque

What I learned

Agentic systems fail in the seams: tool schemas, timeouts, and incomplete tickets. Tracing wasn't optional; it was how the product became trustworthy. Routing across multiple models helped cost and latency, but a single clear orchestration layer mattered more than which model sat behind each step.

Stack

Go, PostgreSQL, Gemini, Groq, DeepSeek, LangSmith, Telegram Bot API.