Advanced Recommendation Engine

Recommendation systems are the backbone of user engagement in consumer-facing products. This project implements a dual recommender architecture — content-based filtering combined with collaborative filtering — deployed on Kubernetes at production scale.

Architecture

The system has two parallel recommendation pathways that are blended at serving time:

Content-Based Pathway

Item features (cuisine type, price range, location, rating) are encoded into dense embeddings. For cold-start items with no interaction history, this pathway provides reasonable recommendations based on item metadata alone.

Collaborative Filtering Pathway

User-item interaction data flows through a TensorFlow-Recommenders model trained on implicit feedback (clicks, bookmarks, bookings). The model learns user and item embeddings in a shared vector space where cosine similarity predicts preference.

Graph Embeddings with Cleora

Beyond the standard matrix factorization approach, I incorporated Cleora graph embeddings to capture higher-order relationships. Items frequently co-visited or co-booked form edges in a bipartite graph, and Cleora produces embeddings that encode this structural information. This proved especially valuable for serendipitous recommendations — surfacing items users wouldn’t find through simple similarity matching.

Serving Infrastructure

Model serving uses KServe on Kubernetes, providing:

• Autoscaling: Scales from 2 to 20 replicas based on request throughput
• Canary deployments: New model versions roll out to 10% of traffic first
• A/B testing framework: Configurable traffic splits for offline evaluation of model variants
• Latency: P99 inference latency under 100ms at peak load

The serving pipeline processes 1M+ daily events with 2M+ monthly active users.

Results

Metric	Before	After
Click-through rate	Baseline	+20%
Booking conversion	Baseline	+15%
Cold-start coverage	40%	78%

The graph embedding component was responsible for most of the cold-start improvement — items with fewer than 10 interactions saw a 2x improvement in recommendation quality.