Case Study · My role: AI Engineer · 2026
From business idea to legally registered company.
An AI platform that turns “I have a business idea, what now?” into a personalized roadmap for Flemish starters. One conversational coach, one visual 8-phase plan, grounded in verified sources.
01 · The Challenge
First-time founders face a wall of fragmented legal and administrative information.
Every step lives on a different government website, so most starters pay an accountant to translate it or give up. And the hard constraint shaping everything: hallucinated legal advice is a liability event, so the AI architecture had to make it structurally impossible.
02 · The AI Stack I Built
Three AI surfaces, each with its own constraints.
Intake Agent
Claude Sonnet 4.6 + tool use
Interviews the founder one question at a time, filling a structured intake via tool use. Questions adapt per sector; a postcode gate blocks out-of-scope users before any LLM cost.
Advisor Agent
Claude Sonnet 4.6 + RAG
Coach scoped to the user's current phase and task. Every message triggers pgvector retrieval; web search activates only when retrieval is weak, and never for legal queries.
Query Classifier
Claude Haiku 4.5
Routes every message by category. Anything possibly legal is classified legal, where web search is always off. About $0.001 per message for a hard safety guarantee.
03 · The 7-Agent Legal Engine
The legal phases aren't generated by an LLM at all. That's by design.
The two legal phases come from a deterministic pipeline of 7 agents: every legal item is traceable, reviewed by our legal lead, and can't hallucinate.
04 · LLM Safety
Safety rules enforced in code, not in hope.
Prompt injection defense
User input is never injected into system prompts. All user data is wrapped in XML tags with explicit treat-as-data instructions.
Fail-closed rate limiting
If Redis is down in production, requests get a 503 instead of unlimited passthrough, so an outage can't turn into a five-figure API bill.
05 · Stack