Maisha Design Platform

Two years as the engineering function for an interior design firm, covering both the machine learning and the unglamorous software that makes a business run.

Generative AI in the design loop

Generative AI and conditional GANs applied to the creative process, generating and iterating on design options far faster than by hand. Conditional generation was the right tool because the output has to respect constraints: a room, a style, a palette. Unconditioned generation is a novelty.

The platform

• A full-stack application centralising furniture offerings from many suppliers into one database, replacing a scattered pile of supplier PDFs and websites.
• A project-management frontend that generates invoices and lets designers browse catalogues while specifying a project.
• Automated scrapers keeping the catalogue current without manual re-entry.
• LLM-enhanced category matching, to reconcile the fact that every supplier names and categorises the same thing differently. That kind of messy semantic matching was impractical before language models and is nearly free now.

30% reduction in project turnaround time, sustained across the business rather than measured on a demo.