AI System
Certification Validator
Automated TÜV/ABE certificate verification for a German tire retailer — reducing return risk and removing the compliance bottleneck.
20-60s per check
~$0.004 per check
DE market
Overview
In Germany, selling aftermarket wheels requires verifying compliance with TÜV/ABE certification documents — dense PDFs with tables, restriction codes, and cross-references. If a retailer ships non-compliant parts, the customer has the legal right to return them. For reifen-anton.de, this verification was a manual bottleneck: 5–15 minutes per order, done by the owner personally. This system reads the certificates and validates fitment in 20–60 seconds, cutting return risk and removing the compliance bottleneck.
Problems Solved
- Regulatory complexity: German market requires TÜV/ABE certificate verification for every aftermarket wheel sale
- Return risk: non-compliant sales lead to returns, refunds, and potential liability
- Manual bottleneck: checks took 5–15 min per order, only the owner could do them
- Subtle vehicle differences: similar models differ by fuel type, drivetrain, power — easy to miss manually
- Scaling: couldn’t grow order volume without the owner personally checking every order
Architecture
4-layer system:
- Context Engineering — order parsing, PDF download, synonym enrichment
- AI Agent — multi-turn reasoning with tool-calling (search_certs, read_page, get_full_pdf, return_result)
- Structured Output — validation results with Auflagen, fasteners, hub rings
- Self-improvement — feedback loop, PDF annotation, prompt editing, synonym management
Key Features
- Two input modes: order ID (auto-parse from shop) or direct URL + vehicle params
- Multi-turn reasoning agent with up to 20 turns per check
- 4-step workflow with mandatory 7-point self-check before returning result
- Real-time SSE streaming of reasoning and tool calls
- Extracts: compatibility status, restrictions (Auflagen), wheel specs, fastener details, hub rings
- Self-improvement tools: feedback UI, PDF annotation, prompt editor, synonym management
- All improvements without code changes or redeployment
Results
- Production deployment on Railway
- Full request logging in Supabase with cost tracking
- Eval testing with 8 ground-truth cases
- MCP Server for integration with other AI agents
- Average cost ~$0.004 per validation check
Stack