SleepMatch AI

An explainable AI-assisted mattress consultation that turns sleep needs into transparent recommendations and a smarter retail handoff.

Comment

Inspiration

My name is Mubashir KC. I work in mattress retail in Dubai, where I regularly see customers face one difficult question:

Which mattress is right for me?

A mattress decision is personal, but the showroom conversation can be inconsistent. The answer depends on connected factors such as sleeping position, body comfort, heat, partner movement, support needs, and feel preference. Customers may receive a lot of product information without a clear way to compare options, while staff need a structured handoff to continue the conversation confidently.

SleepMatch AI came from that real showroom problem. I built it to turn an unstructured mattress conversation into a clearer, explainable consultation workflow for both the customer and the retailer.

What it does

SleepMatch AI is an explainable AI-assisted B2B retail consultation and sales-handoff platform. It is not just a consumer mattress quiz.

The customer completes five short steps:

  1. Basic information
  2. Sleep profile
  3. Sleep and body needs
  4. Comfort preferences
  5. Recommendation

A deterministic weighted-rule engine compares the completed profile with a configured retailer catalogue and returns a locked top-three mattress comparison.

Each result shows:

  • weighted-rule alignment percentage
  • matched customer needs
  • relevant product strengths
  • honest trade-offs to test in the showroom

The percentages represent weighted-rule alignment, not statistical confidence, medical prediction, or diagnosis.

Optional add-ons such as pillows, protectors, and toppers can support the sleep setup, but they never change the mattress ranking.

How GPT-5.6 is used

GPT-5.6 is an optional explanation and staff-support layer. It never chooses, scores, reorders, or changes the mattress recommendations.

After an explicit action, GPT-5.6 explains the already locked recommendation in clear customer language. In the protected Retailer Workspace, it can create grounded staff talking points, questions to confirm, cautions, accessory suggestions, and a consent-aware follow-up draft.

GPT-5.6 explains. The deterministic engine ranks.

Retailer workflow

After a consultation, the protected Retailer Workspace gives staff a structured handoff. Staff can review saved leads, filter by customer consent, open a printable deterministic consultation brief, and use a separate GPT-5.6 Retail Sales Copilot.

WhatsApp follow-up is available only when the customer has given consent.

Public demo and privacy

This Build Week submission is a public demonstration of the product foundation.

The live demo uses a fictional retailer and an original demo catalogue. I intentionally did not use real retailer names, mattress brands, product catalogues, or customer data, in order to protect commercial information, brand assets, and customer privacy.

The architecture is designed as a tenant-configurable foundation for future retailer deployments; full multi-tenancy is not claimed as complete in this demo.

GPT is called only after an explicit user action. Customer name, phone number, consent status, exact age, height, weight, IP address, session identifiers, and raw free-text notes are excluded from the GPT payload.

SleepMatch AI does not make medical diagnosis or treatment claims.

How I built it

I built SleepMatch AI with Python, Flask, Jinja templates, HTML, CSS, JavaScript, Supabase/PostgreSQL, Render, and GitHub.

For the AI layer, I used the OpenAI Responses API with Structured Outputs and GPT-5.6. Codex helped me implement, test, debug, and deploy the application, while I led the real retail problem definition, recommendation logic, product decisions, catalogue mapping, and final product direction.

A key challenge was keeping generative AI useful without allowing it to invent or alter recommendations. The final design keeps the deterministic engine as the source of truth, validates the GPT output against the locked recommendation, and preserves deterministic fallbacks.

Demo note

The public demo is hosted on Render’s free tier. After a period of inactivity, the first request may take a short moment to wake up; once active, the consultation flow is available normally.

Built With

Share this project:

Updates