Machine LearningJan 2024 - Mar 2024

Wine Recommendation Engine

Hybrid ML recommendation engine for wine discovery, combining collaborative filtering, content-based filtering, and behavioral signals like order and search history for highly personalized suggestions.

Architecture Flow

data flow · live
User Signals
Orders · Search
Collaborative
User similarity
Content-based
Varietal · Region
Hybrid Fusion
Ranked Wines
Personalised

Key Achievements

  • Built a hybrid recommendation pipeline fusing collaborative filtering and content-based filtering for superior accuracy
  • Incorporated user order history and search history as behavioral signals for intent-aware personalization
  • Modeled wine-specific content attributes including varietal, region, tasting notes, and price for content-based matching
  • Addressed cold-start problem through content-based fallback for new or low-activity users
  • Designed scalable pipeline capable of serving personalized recommendations across a large wine catalog
  • Improved product discoverability by surfacing wines users would not have found through manual browsing

Core Challenge

Wine discovery is highly subjective and catalog-dense — users struggled to find wines matching their taste preferences, and the platform lacked any personalization layer to guide purchasing decisions beyond basic filters.

Solution

Engineered a hybrid ML recommendation engine that combines user-based collaborative filtering with content-based filtering on wine attributes, enriched by real behavioral signals from order and search history to produce personalized, intent-aware wine recommendations.

Timeline
Jan 2024 - Mar 2024
Team
Lead Engineer
Status
Production Ready
Category
Machine Learning
Live Preview View Code

Deep Dive

Designed and developed a domain-specific recommendation engine for a wine platform, built to surface highly relevant wine suggestions tailored to each individual user. Generic recommendation approaches fail in niche domains like wine where personal taste, regional preferences, and varietal nuances play a significant role — so the system was architected as a hybrid engine that combines multiple recommendation strategies to maximize relevance and discovery.

The engine ingests and processes three core behavioral and content signals: a user's order history to understand past preferences and purchase patterns, search history to capture intent and curiosity signals, and product-level wine attributes for content-based matching. These signals feed into a hybrid pipeline that blends Collaborative Filtering — finding wines loved by users with similar taste profiles — with Content-Based Filtering that matches wines based on varietal, region, tasting notes, price range, and other domain-specific attributes.

The fusion of these approaches ensures the engine handles both warm users with rich interaction history and cold-start scenarios where limited data is available, delivering meaningful recommendations across the entire user base.

Tangible Impact

Delivered a fully operational personalized recommendation system that increases product discoverability, drives repeat purchases through taste-matched suggestions, and handles both new and returning users effectively through its hybrid architecture.

Tech Stack

PythonScikit-learnPandasNumPyCollaborative FilteringTF-IDFFastAPIPostgreSQL

© 2024 NIKHIL

BACK TO TOP ↑