Multimodal Real-Estate Publishing Engine

The problem

Publishing a property listing requires more than writing a description. Images, existing property data, domain terminology, structured metadata and downstream publishing requirements all need to agree. At Co-libry, I helped turn that fragmented process into a multi-step AI pipeline.

Images as the starting point

The pipeline used property imagery as a rich source of information. Image-processing and multimodal steps extracted useful signals, while existing CMS and ERP data supplied facts that should not be guessed from pixels alone.

Extraction, enrichment and grounding

Python workflows combined LLM prompting, structured tagging and retrieval-augmented generation. Domain-specific information was retrieved to ground the generated content, and intermediate outputs were represented in structured form so later steps could validate and reuse them.

A staged publishing workflow

• Ingest property images and available source data.
• Extract visual and structured attributes.
• Enrich the context with relevant domain knowledge.
• Generate descriptions and listing metadata.
• Validate outputs before passing them into CMS and ERP publishing flows.

Why the architecture mattered

Breaking the task into explicit stages made the system easier to inspect and improve than a single opaque prompt. It also created clear places for grounding, validation and integration, patterns that later became central to production agentic systems.

What I learned

The project taught me to treat model output as one component in a wider data product. The quality of the final listing depended on source quality, orchestration, validation and system integration just as much as it depended on the model.