SOZOColour: Search by Seeing

Turn what customers see and feel into products they can buy.

Comment

The problem

Colour-led retailers have a search problem: customers often recognise the shade they want but cannot name it.

A shopper may want the coral from a sunset, the green from a hotel lobby, or a nude that complements their skin, yet conventional search expects words such as terracotta, sage, warm, cool, muted, or vivid. When the customer does not know the catalogue vocabulary, filters and keyword search break down.

SOZOColour began with this real problem at Nail Deck. With 806 product colours, customers regularly knew what they wanted when they saw it but struggled to verbalise it well enough to find the right shade.

The solution

SOZOColour is an AI colour-reasoning layer for colour-led commerce. It translates visual references and subjective human intent into genuine products from a merchant's catalogue.

The shopper-facing promise is simple:

Find the colour you saw—or the feeling you meant.

A user can upload an image, tap the colour that caught their attention, and receive the closest products from a real catalogue. They can then refine the result using natural language such as:

  • “A Bali sunset, but sadder.”
  • “Keep this warmth, but make it more sophisticated.”
  • “I like this nude, but I want something softer.”

The experience deliberately separates two different jobs:

  • Closest to what you saw — an objective match calculated with deterministic colour science.
  • Closest to what you meant — intent-aware recommendations grounded in the user's language, mood, context, and desired perceptual shift.

Why this is a Work & Productivity product

The interface is used by shoppers, but the product is built for merchants.

A colour-led retailer can place SOZOColour above its existing catalogue as a storefront component, white-label discovery tool, or future API. The merchant provides the products; SOZOColour helps customers express otherwise unsearchable intent and turns it into a useful shortlist of items they can actually buy.

Nail Deck is the first working deployment and catalogue. The same approach can extend to:

  • interior paint and wallpaper;
  • cosmetics and hair colour;
  • fashion, footwear, and accessories;
  • furniture and home décor;
  • stationery, textiles, ceramics, and materials.

Rather than asking whether everyone wants to study colour, SOZOColour addresses the moments when colour is already central to a purchasing or design decision.

How the experience works

  1. The user uploads a photograph or selects a built-in sample image.
  2. They click or tap a point in the image to choose a reference colour.
  3. A deterministic matching engine compares that colour with genuine catalogue records and identifies the closest objective matches.
  4. The user may add a feeling, setting, memory, or requested transformation.
  5. GPT-5.6 interprets subjective intent such as warmer, quieter, more playful, more luxurious, or less corporate.
  6. GPT-5.6 reasons only over a controlled set of real catalogue candidates rather than inventing products.
  7. The result is presented as a concise Colour Story, clearly distinguishing visual similarity from perceptual fit and explaining each recommendation.

The goal is not to promise pixel-perfect physical matching across every screen, camera, lighting condition, and material. SOZOColour is a discovery and inspiration layer that helps people reach a much better product shortlist before making a final physical selection.

Why it is different

Most colour tools return a hex value, a nearest swatch, or a visually similar item. General visual-search systems can recognise objects and styles, but colour is usually only one feature among many.

SOZOColour combines:

  • perceptual colour measurement;
  • a merchant's canonical catalogue;
  • contextual visual references;
  • semantic interpretation of ambiguous human language;
  • explainable, catalogue-grounded recommendations.

The differentiation is not simply matching an orange photograph to orange products. It is understanding transformations such as:

“Keep the warmth of this sunset, but make it quieter and more reflective.”

That kind of intent has no direct RGB value and cannot be expressed reliably through a conventional filter.

Technical implementation

The Build Week prototype uses a shared colour engine and a server-side AI boundary:

  • image and manual colour selection in the browser;
  • perceptual catalogue matching using Lab colour values and CIEDE2000 distance;
  • a canonical Nail Deck catalogue containing 806 genuine products;
  • a dependency-light Node server that keeps OPENAI_API_KEY out of the browser;
  • a secure /api/colour-story route;
  • GPT-5.6 through the OpenAI Responses API;
  • structured model output with server-side validation of every returned product ID;
  • resilient deterministic results when an AI request is unavailable;
  • responsive shopper-facing results that separate objective and intent-aware recommendations.

Deterministic colour science handles measurable similarity. GPT-5.6 handles the part traditional filters cannot: translating perception, emotion, and ambiguous language into a useful product-discovery decision.

How Codex was used

Codex is the primary development collaborator for this project. It audited and preserved the pre-existing SOZOColour application, identified the shortest reliable path to a complete product, refactored shared colour logic, created the secure server boundary, implemented the GPT-5.6 workflow, developed the new result experience, added fallback states and tests, verified the runtime, and documented the final architecture and decisions.

Keeping the work in one primary Codex thread also provides a traceable record of the core implementation, product trade-offs, testing, and iteration.

What existed before Build Week

The project already included:

  • the SOZOColour sphere and manual exploration interface;
  • image upload and pixel sampling;
  • Lab-based catalogue matching;
  • a real 806-product Nail Deck catalogue;
  • product links, responsive layouts, and a separate creator QA mode.

What was added for Build Week

The focused Build Week extension turns those components into an end-to-end merchant product-discovery experience through:

  • a one-click sample journey for judges;
  • a natural-language intent composer;
  • production-grade CIEDE2000 matching shared by browser and server;
  • a secure server-side OpenAI route;
  • GPT-5.6 structured colour reasoning over constrained catalogue candidates;
  • clear “what you saw” versus “what you meant” result groups;
  • validated, explainable recommendations using genuine product IDs;
  • loading, failure, and deterministic fallback states;
  • focused tests, launch tooling, and Build Week documentation.

Commercial direction

The initial commercial model is B2B2C: merchants deploy SOZOColour, while shoppers use the embedded discovery experience. Potential delivery models include a commerce-platform app, white-label widget, usage-based API, or enterprise catalogue integration.

Future extensions may include merchant analytics showing what customers were trying to express, guided colour profiling, instinctive palette swiping to learn preferences, downloadable showroom reports, and personalised discovery across multiple colour-led industries.

SOZOColour's broader ambition is to make perceptual intent searchable—and make every colour-led catalogue easier for humans to navigate.

Built With

Share this project:

Updates