RugDesk

Handmade rug inventory is unusually messy: every item is unique, dimensions vary by inches, names are subjective, and availability changes. A fluent recommendation is still wrong if it points to an unavailable item, the wrong store, or a rug that cannot fit.

The key design question became: what should the model be allowed to consider in the first place?

I built the operational and customer-facing sides together so the recommendation experience was grounded in the same source of truth staff used to manage inventory.

• Photo-first inventory intake with AI-assisted enrichment and staff review.
• Tenant-scoped search, saved items, and staff handoff workflows.
• A customer-facing storefront and conversational product discovery experience.
• Structured session state for follow-up requests such as larger, warmer, or more like the first option.

The recommendation pipeline first resolves the store, loads structured session state, filters for valid inventory, applies hard constraints and deterministic pre-ranking, and only then gives a small candidate set to the model.

Model-returned IDs are checked against that candidate set before the UI can render them. Invalid references are removed, no-match is a supported outcome, and fallback paths preserve a useful product when a model or provider fails.

• Rotation-aware dimension matching with tolerance windows.
• Hard constraints separated from soft ranking preferences.
• Structured model outputs with schema validation.
• Tenant isolation enforced server-side rather than trusted to prompts.

I concluded RugDesk in July 2026 after taking it from an ambiguous retail problem to a deployed, end-to-end product. The project sharpened how I reason about multi-tenant data, model boundaries, fallbacks, and the difference between a convincing demo and a reliable product.

The archived demonstration uses sample inventory. Transactions, customer communication, and staff mutations are disabled.