Art Matched to Your Space

For this project at LUMAS, I built a feature that recommends curated artworks based on a photo of the customer's own room. Upload an image, receive a personal selection — each recommendation accompanied by a written explanation of why that particular piece belongs in that space.

At the core of the pipeline sits a multimodal architecture: a vision encoder extracts image embeddings from the room photograph, which are then passed — alongside textual descriptions of the artwork — into a language model for joint reasoning. Multiple models work in concert across the pipeline, and the real craft lies in their orchestration: structuring the handoffs, shaping the inputs, and ensuring the outputs stay coherent and genuinely useful rather than just plausible-sounding.

The result speaks for itself — recommendations that feel considered, not generated.