<aside> 💡

A three-signal purchase intelligence system built for Phia's personalisation roadmap

</aside>

</aside>

Phia Wardrobes Intelligence is a project that makes intelligent shopping personalisation a reality. It understands user taste based on your collection and aesthetic inspiration for a deeply personal shopping experience, while intelligently working with you to style you based on today’s needs.

<aside> 👩🏻‍💻

My Github!

VeedhiBhanushali - Overview

</aside>

Project Repo

GitHub - VeedhiBhanushali/Phia-wardrobe-intelligence

</aside>

Overview

The question Phia answers: Is this the best price for this item?
The question this project answers: **Should I buy this, given who I am, what I already own, and what I'm looking for right now?**

The Problem

Cold-start personalisation and real-time intent inference are the two hardest problems. 50%+ return rate reduction for brand partners depends on solving both.

Today, every Phia user sees the same feed. Personalisation requires purchase history that doesn't exist yet. And even when history accumulates, most systems conflate two very different signals: who you are aesthetically, and what you need right now. A Quiet Luxury shopper browsing for a ski trip shouldn't get more cashmere.

This project is the working proof of concept for the personalisation layer that closes both gaps and adds a third: predicting whether a purchase will integrate with what the user already owns, before they buy it.

The Core Idea

Every recommendation in this system is the intersection of three signals, each distinct:

Taste     — persistent    — who you are aesthetically
Wardrobe  — persistent    — what you already own
Intent    — ephemeral     — what you need right now

Most systems have one signal (taste, via purchase history) or none (popularity ranking). Having all three, and knowing when to weight each, is what makes the difference between a feed that feels generic and one that feels like it knows you.

Three capabilities. Three direct mappings to Phia's Series A thesis.

Cold-start personalisation that produces meaningful recommendations before any purchase history — via visual aesthetic extraction from Pinterest boards using a fashion-domain embedding model. Phia's stated cold-start problem, solved at demo scale with measurable results.

Real-time intent inference that treats long-term taste and present-moment need as separate signals, via session browsing coherence that modulates ranking without overriding taste. A Quiet Luxury user shopping for festival outfits gets festival outfits. Phia's intent problem, implemented and measured.

Purchase quality prediction as a return-rate proxy, via a wardrobe-integrated confidence score that evaluates whether an item will actually work with what the user already owns. Phia's commercial claim is 50% return rate reduction. This is the system that would produce it.

Demo

Phia Wardrobes & Aesthetic Analysis

https://youtube.com/shorts/W06X1XKjkM8?feature=share

Ask Phia Agentic Stylist Chat