AWS re:Invent 2025 - Slickdeals' AI-powered deal discovery with Amazon Sagemaker (SMB206)

Summary of AWS re:Invent 2025 - Slickdeals' AI-powered Deal Discovery with Amazon SageMaker

Overview of AI and Machine Learning

• Artificial intelligence (AI) is the ability to replicate tasks requiring human intelligence
• Machine learning (ML) uses large datasets to make probabilistic decisions with high certainty
• Subsets of AI include classification, predictive, and generative AI powered by large language models

Introduction to Amazon SageMaker

• Amazon SageMaker is a comprehensive ML service that enables business analysts, MLOps engineers, and data scientists to build, train, and deploy models on AWS
• SageMaker provides various tools to accomplish ML tasks regardless of ML expertise

Slickdeals' Business Challenges

• Scaling data and infrastructure as the business grew, including limitations in processing impressions data
• Lack of real-time interaction data, making it difficult to effectively run personalization
• Inability to understand deal performance quickly enough before deals expired

Slickdeals' Modernization Approach

1. Migrated data platform to Databricks to enable a more scalable, pipeline-based system
2. Migrated applications from on-premises LAMP stack to EKS orchestration
3. Enhanced data collection instrumentation to support real-time processing
4. Brought personalization and search capabilities in-house to leverage their own data and insights

Personalization and Deal Scoring Frameworks

• Personalization uses a two-stage process:
  1. Retrieval stage with Siamese model to find similar deals based on text embeddings
  2. Filtering and ranking stage using heuristics and XGBoost model
• Deal scoring framework has two components:
  1. Engagement score based on user interactions (votes, comments, clicks, etc.)
  2. Content score based on deal attributes (category, discount, brand, etc.)
• These scores are used to power personalization, operational tools, quality assurance, and deal sourcing

Technical Architecture

• Real-time instrumentation platform using API Gateway, Lambda, and Kafka to capture events
• Databricks used for data processing and powering personalization and deal scoring
• Elasticsearch and SageMaker used for search and query classification

Business Outcomes

• Reduced personalization and A/B testing turnaround time from 6 weeks to 1 week
• Decreased deal scoring latency from 3 hours to 30 seconds, enabling faster operational decisions
• Achieved a 7% increase in merchant outclicks and revenue from personalization improvements

Key Takeaways

1. Focus on scalability and reducing infrastructure overhead to enable more innovation
2. Leverage your business differentiators and unique data to drive personalization and optimization
3. Build a strong data foundation to enable real-time decision-making and short feedback loops
4. Modernization can unlock innovation and strengthen your ability to differentiate