AI catalog · live
Tiles & Surfaces · Catalog AI
A surfaces and tiles distributor processing supplier catalogs from multiple brands - parsed into clean structured data with AI metadata enrichment.
01The context
Catalogs in PDFs. Data needed in structured form.
The client received supplier catalogs as multi-brand PDFs - text, tables and high-resolution product images all mixed. Manually extracting product data and matching it to inventory was slow and error-prone. The volume was too high for human-only processing.
02The diagnosis
Hybrid AI pipeline - cost-aware, accuracy-led.
Cloud vision APIs would solve accuracy but blow the unit economics. Local models would handle the routine 70% but lacked accuracy at the edge. The right architecture was hybrid - local-first, cloud-on-low-confidence.
03What we installed
The interventions, in order.
- Unstructured PDF parsing — Text, tables, images extracted to clean structured JSON.
- 300 DPI image extraction — Pages rendered at high resolution, deduplicated, matched to products by position.
- Metadata enrichment — Finish, thickness, temperature, texture, style auto-populated from images.
- Hybrid local + cloud vision — Local model handles routine work. Cloud vision called only when local confidence is low.
04The outcomes
Numbers, not deliverables.
300DPIImage extraction quality
LiveIn production
HybridCost-optimised inference
06What is live today
Still running
What runs after we step back
- PDF-to-JSON pipeline — Live across multi-brand supplier catalogs.
- Hybrid inference — Local-first routing with cloud fallback.
- Metadata model — Auto-enrichment running on every new product.
Your story
Two hours. One conversation. On us.
Bring the question. We bring twelve years of pattern recognition.