AI-Powered Fashion Discovery

Discover fashion that
understands you

A TikTok-style shopping feed powered by real-time deep learning. Every swipe, like, and second spent trains a model that learns your taste.

Acne Studios

Oversized Wool Coat

$890

Lemaire

Twisted Silk Blouse

$620

Toteme

Draped Knit Dress

$450

About Looksy

Fashion discovery, reimagined

Online fashion shopping is broken. Consumers face endless scrolling through generic product grids, irrelevant search results, and recommendation systems that haven't evolved past “customers also bought.” The result: decision fatigue, missed discovery, and a widening gap between what shoppers want and what they find.

Looksy solves this with AI-native product discovery. We replace the traditional product grid with a personalised, TikTok-style feed that learns from every interaction — every swipe, every pause, every like. Our real-time deep learning engine understands not just what you search for, but what you respond to visually and contextually.

The Problem

Fashion e-commerce relies on static category browsing and keyword search. Consumers spend more time searching than discovering. Conversion rates stagnate because product discovery hasn't kept pace with content discovery.

Our Solution

An AI-powered discovery feed that mirrors how people consume content — not how retailers organize inventory. Real-time personalisation powered by deep learning, multi-modal search combining text, visual, and semantic understanding.

Target Audience

Fashion-forward consumers aged 18-35 who shop across multiple retailers. Digital natives who expect TikTok-level personalisation from their shopping experience. Currently focused on the UK market with 27 integrated retailers.

Live Demo

Experience it yourself

Every swipe trains the model. Like products to see your recommendations improve in real-time.

How It Works

Intelligence in every swipe

01

Browse

Swipe through a personalized feed of products from 27+ UK fashion retailers. Each product is curated by our recommendation engine.

02

AI Learns

Every interaction — likes, dwell time, images viewed — feeds our real-time ML model. Six distinct signals train the recommendation engine.

03

Feed Adapts

Within seconds, your feed evolves. Real-time parameter sync means no waiting for overnight batch retraining. Your taste profile updates continuously.

Technology

The engine behind the feed

A production-grade ML system combining real-time deep learning, multi-modal search, and automated data pipelines.

Recommendation Engine

Learns in real-time

Our two-tower neural architecture separates user preference modeling from product understanding, enabling sub-100ms inference at scale.

Unlike traditional batch systems, our Monolith-based engine uses real-time parameter sync — user preferences propagate to the serving model in seconds, not hours.

<100ms inference latency
User Interaction
Kafka Event Stream
Monolith Two-Tower Training
Parameter Sync
Updated Embeddings → Better Recs

Search Engine

Three search methods, one query

Search “red summer dress” and our system understands both the text meaning and the visual aesthetic.

Reciprocal Rank Fusion combines results from semantic text search (768-dim Marqo embeddings), visual search (512-dim CLIP embeddings), and keyword matching — ensuring no single method dominates.

100K+ products in <50ms
Semantic · 768-dim
Visual · 512-dim
Keyword · BM25
Reciprocal Rank Fusion
Ranked Results

Data Pipeline

From retailer to recommendation

Automated crawling and parsing of 27 UK fashion retailers. AI-generated CSS selectors via Gemini 2.0 ensure robust extraction across diverse site architectures.

Every product generates three embedding representations — CLIP for visual understanding, Marqo for semantic text, and learned embeddings from our two-tower model — all indexed in PostgreSQL with pgvector.

27 retailers, automated daily
27 Retailer Sites
Cloud Run Crawlers
AI Parser · Gemini 2.0
CLIP
Marqo
Learned
PostgreSQL + pgvector

Architecture

Full-stack intelligence

From user interaction to model update in under a second. Every layer purpose-built for real-time personalization.

User · iOS / Web

Next.js + Capacitor

Interactions ↓
Queries ↓

Recommendation API

FastAPI · Python

Hybrid Search

RRF Fusion Engine

PostgreSQL + pgvector

100K+ products · 3 embedding spaces

Kafka

Event Streaming

Monolith

Online Training

Firebase

Auth · Firestore

0K+

Products

0

Retailers

<100ms

Latency

0

Embedding Models

Real-time

Online Learning

ML Framework

Built on Monolith

Our recommendation engine is powered by Monolith — the open-source deep learning framework developed by ByteDance, battle-tested at TikTok scale for billions of users.

Read the paper →

Collisionless Embedding Tables

Unique representations for every product and user. Cuckoo hashing guarantees O(1) lookups with zero hash collisions across millions of entities.

Real-Time Parameter Sync

Embedding updates propagate from training to serving in under 100ms — without reloading the model. Only changed parameters are synced via gRPC.

Two-Tower Architecture

Separate user and item towers enable independent scaling. User preferences are computed at request time; product embeddings are pre-computed and indexed via pgvector.

Team

Built by engineers

ZP

Zak Pyemont

Founder

Backend engineer at Spotify, building ETL and transcription pipelines at scale. Looksy combines deep experience in production data systems with a passion for applying real-time ML to consumer products.

LinkedIn
ML SystemsData PipelinesPythonTensorFlowGCPReal-time Systems

Looksy is a solo-founded startup. Every line of code — from the ML recommendation engine to the iOS app — is built by one engineer.