How to Build a SaaS Platform Support Copilot ?

Most support bots guess. This one can’t. In five minutes: how to build a read-only AI support copilot that answers only from evidence and always cites its sources. The test subject is Loopline, a fictional ticketing app with a seeded notification bug. You’ll see typed and validated model output, routing, safe tools, a bounded agent loop, hybrid retrieval (RAG), a call graph, memory, planning, permission-aware search, refusal of destructive commands, and prompt-injection defense — with real code and real test output from the repository.

This is a condensed summary of the core build (lessons 00–13 of the 19-lesson course). The full step-by-step lessons, code and tests are in the repository:
https://github.com/scheffershen/saas-support-copilot

Timestamps:

• 0:00 The problem: a support copilot that can’t guess
• 0:52 The typed brain: structured output & routing
• 1:58 Tools and the agent loop
• 2:59 Finding knowledge: RAG, call graph, memory, planning
• 3:56 Safety: permissions, refusal, prompt injection
• 4:59 Recap and where to find the code

Lessons covered (repository folder `lessons/`):

• 00 What we’re building
• 01 Python project foundations
• 02 The smallest useful LLM call
• 03 System prompts and structured output
• 04 Decisions, routing, and workflows
• 05 Tools and safe tool contracts
• 06 The agent loop
• 07 Conversation state and memory
• 08 RAG fundamentals
• 09 Knowledge graphs for code relationships
• 10 Permission-aware search
• 11 Planning feature feasibility
• 12 Human approval and refusal
• 13 Prompt injection and data-exfiltration defense

Learning prototype, not a production-security guarantee: the repository’s production-readiness checklist lists what is still missing. Loopline is fictional; no real company data is used.

The narration is an AI-generated version of the author’s own voice.

#AIAgents #RAG #LLM #Python #PromptInjection #SoftwareEngineering #SaaS #Tutorial