ProjectsFlavours Collective

— BUILD

Work in Progress

FlavoursCollective

An intelligent product discovery and curation platform built to bridge the gap between online shoppers and the products they love. Currently in active development as a full application.

— OVERVIEW

Rather than relying on manual browsing across dozens of webshops, Flavours Collective automates the process of product discovery, aggregation, and organisation, giving users a single curated space to track items from anywhere on the web.

The platform combines modern web scraping infrastructure, a microservices architecture, and on-device AI inference to deliver a seamless experience that is fast, private, and extensible.

— THE PROBLEM

Online shopping is fragmented. A consumer interested in a particular aesthetic, say Scandinavian furniture or independent streetwear, must visit tens of individual webshops, bookmark pages manually, and repeatedly return to check availability and price. There is no neutral, cross-retailer layer for discovery.

Existing aggregators either lock users into walled gardens, rely on retailers to opt in through affiliate programmes, or surface results heavily influenced by advertising spend rather than genuine relevance. The result is a discovery experience that serves retailers more than it serves shoppers.

— THE SOLUTION

Flavours Collective takes a fundamentally different approach. Instead of waiting for retailers to share their catalogue, the platform actively and intelligently extracts product data directly from the source at the moment the user requests it.

Users create Boards, themed collections of products from any webshop they choose. A Board might represent a mood, a project, a budget, or a gift list. Products are pulled into those boards on demand, kept up to date, and made shareable with others.

— ARCHITECTURE

01

Intelligent Scraping Engine

A Python-based scraping service capable of navigating modern, JavaScript-rendered webshops. Rather than relying on brittle hand-written selectors, the engine uses a dynamic discovery process to identify product fields without prior configuration. It can ingest products from webshops it has never seen before.

02

Orchestration Layer

A Java-based orchestration service that sits between the frontend and the scraping engine. It handles user authentication, board management, data persistence, and coordinates requests across services. All scraped data flows through here before being stored and associated with the correct user and board.

03

Frontend and Browser Extension

Users interact through a responsive web application. A companion Chrome Extension allows users to trigger scraping directly from the retailer's website. Browse a webshop, click the extension, and products are added to whichever board the user selects. No copy-pasting required.

— AI AND ML

Current: LLM-Assisted Extraction

When the scraping engine encounters an unfamiliar page structure, it consults a locally running large language model to reason about the HTML and generate a reliable extraction strategy. The model runs entirely on-device. No data leaves the user's infrastructure and no third-party API calls are required.

Planned: Fine-Tuned Qwen 3B

The next phase involves fine-tuning Qwen 3B, a compact open-source language model, on a curated dataset of webshop HTML and corresponding product schemas. The goal is a purpose-built extraction model that understands the vocabulary of e-commerce HTML, runs faster than a general-purpose model, and generalises reliably to new retailers. MLflow is integrated throughout to track experiments, log metrics, and manage model versions as the fine-tuning work progresses.

— TECHNOLOGY HIGHLIGHTS

Microservices Architecture

Each concern is isolated into its own service, independently deployable and scalable.

Stateless API Design

All services communicate over HTTP with JWT-based authentication, making the system horizontally scalable.

Browser-Grade Rendering

The scraping engine uses a real browser under the hood, handling JavaScript-heavy storefronts that defeat traditional scrapers.

AI Inference

LLM inference runs SLM, keeping user data private and costs predictable by hosting the model ourselves.

Experiment Tracking

MLflow provides a full audit trail of model experiments and scraping performance across retailers.

Fine-Tuned Qwen 3B (Planned)

A purpose-built extraction model trained on e-commerce HTML, replacing general-purpose prompting with a fast, specialised local model.

— USE CASES

The Collector

Someone curating a wishlist across multiple independent brands. They use the extension to pull products into a single board, share it with friends, and revisit it when a sale hits.

The Designer

A creative professional building a mood board of products from around the web for a client project. Boards are organised by theme and can be viewed by collaborators.

The Bargain Hunter

A user who tracks a specific product across multiple retailers, checking back regularly as the platform refreshes product data on demand.

— ROADMAP

  • 01Fine-tuned Qwen 3B extraction model, replacing prompt-based extraction with a specialised, fast, local model
  • 02Price history tracking, surfacing price trends over time for products in a user's boards
  • 03Availability alerts, notifying users when an out-of-stock item becomes available
  • 04Public board discovery, allowing users to follow and explore boards curated by others, creating a social discovery layer
  • 05Retailer analytics, aggregate anonymised insights into product trends across the catalogues the platform has indexed

— CLOSING

Flavours Collective is being built with the conviction that product discovery should be open, intelligent, and user-first. By combining robust web infrastructure with on-device AI, the platform aims to give individuals the kind of aggregation and curation power that has historically only been available to large retailers and affiliate networks.

The project is currently in active development. If you are interested in following its progress or discussing collaboration, get in touch.

Get in Touch