← Deployment Log

Case study: premium DACH e-commerce

Anonymised PIM and AI content engine project for a premium Vienna shop. Timeline: April–May 2026.

1. Faza 1

   Reconnaissance — As-Is architecture

   SSOT in Google Sheets ([PRODUKTE], [MARKEN], [AI_TUNER]). No-code front (AppSheet). Description generation via simple prompt concatenation in Cloud Functions. Problem: global persona clashed with B2B brand requirements — hallucinations and broken tone (Attention Drift).

2. Faza 2

   Technology pivot

   AppSheet dropped for modular Streamlit app on GCP Cloud Run. Deterministic Cascade Compiler — surgical JSON tree operation before Vertex AI, separating business and visual layers.

3. Faza 3

   To-Be architecture

   Python (main.py, app.py), scale-to-zero, Local Vision Sorter for 5000+ images. AGARD v4.3 validating 7 frontend sections. Anti-Poetry Shield and Regex scanner block faulty publishes.

4. Faza 4

   Business outcome

   Stable production system on Cloud Run. Reduced information noise in descriptions. Consistent tone per brand without hallucinations. Description creation time cut by an order of magnitude via RAG and cascade compiler.