Upcoming Event: The AI Agent Implementation Framework
Upcoming Event: The AI Agent Implementation Framework
Madison Reed
San Francisco, CA
B2C - Beauty | Hair Color
Amazon SageMaker · Amazon Bedrock · AWS Lambda · Amazon DynamoDB · Amazon S3 · Amazon ECS Fargate · Amazon ECR · Application Load Balancer · Amazon CloudWatch · AWS IAM
Madison Reed, a leading direct-to-consumer hair color company, relies on an interactive online quiz to recommend the right shade to each customer, but that quiz was powered by a manually maintained, rules-based decision tree that could not learn from real customer behavior or scale to new product lines. Avahi, a premier-tier AWS partner, designed and built a machine learning recommendation system on AWS that learns from historical customer data to predict the ideal shade. The solution compared four modeling approaches, deployed the best performer as a real-time Amazon SageMaker endpoint, and wrapped it in an interactive prototype for live testing and feedback. The resulting model placed the correct shade within its top five recommendations 92.9% of the time while still honoring Madison Reed’s expert business rules, giving the company an intelligent, adaptable foundation to replace manual rule-keeping.
Madison Reed is a direct-to-consumer hair color brand known for healthy, salon-quality products and a highly personalized customer experience. A central part of that experience is an online quiz that captures each customer’s hair profile (current color, gray coverage, desired level, and more) and recommends the most suitable shade from an extensive product catalog. As the brand has grown and expanded into new product lines, delivering accurate, scalable recommendations has become increasingly central to its digital commerce strategy.
Madison Reed’s shade recommendations were generated by a rules-based decision tree that had to be hand-tuned by in-house color experts. Every new product line or catalog change required manual reconfiguration and additional quiz logic, making the system slow to adapt and difficult to scale. Because the rules were manually defined rather than learned, they could not improve from the rich signal in customers’ historical quiz responses, purchases, and satisfaction data.
This manual approach created a growing operational burden and a strategic ceiling: as the catalog expanded, maintaining recommendation quality by hand became less sustainable, and the brand remained dependent on third-party tooling with associated cost and lock-in concerns. Madison Reed needed a recommendation engine that could learn from its own data, adapt to new shades automatically, and still respect the hard constraints its colorists rely on, such as never recommending a shade more than two levels lighter than a customer’s current color.
Madison Reed wanted to bring its AI and machine learning capabilities in-house on AWS, reducing reliance on external AI vendors and gaining greater control over cost, data, and the long-term direction of the system. AWS provided a single platform to consolidate data, train and host models, and serve recommendations in real time, and AWS funding programs helped make the initiative possible. The team deliberately chose a traditional and deep learning approach on AWS, rather than a more experimental agentic design, for its predictability, control, and cost efficiency, with Amazon Bedrock available to validate model results.
As a premier-tier AWS partner with deep AI/ML expertise, Avahi brought the data science and GenAI solution architecture talent needed to take Madison Reed from concept to a working, deployable model. Beyond delivery, Avahi provided enablement for Madison Reed’s team, including immersion days and workshops on Amazon Bedrock, and was able to leverage AWS funding programs to reduce the customer’s investment. Avahi owned the engagement end to end: scoping the work, leading the data assessment, developing and rigorously evaluating the models, and delivering full technical documentation and a knowledge-transfer handoff so Madison Reed’s team could operate and extend the system.
Avahi built a serverless, fully managed hair-shade recommendation system on AWS. Working from roughly 82,000 usable records of historical customer data spanning an 89-shade catalog, the team engineered features from each customer’s hair profile and trained and compared four distinct modeling approaches: XGBoost, CatBoost, a deep multi-layer perceptron, and a Hybrid DeepFM architecture combining factorization machines, a deep residual network, and content-based matching for shades with little training history. After a head-to-head evaluation, the Hybrid DeepFM model was selected for production as the best overall performer.
The selected model was deployed on Amazon SageMaker, which handled both model training and real-time inference endpoints returning the top five ranked shade recommendations with confidence scores. AWS Lambda provided serverless orchestration, routing quiz inputs to the model and persisting results, while Amazon DynamoDB stored predictions, user feedback, and session interactions. Amazon S3 held model artifacts and prepared datasets. Customer-facing testing was delivered through an interactive prototype: a conversational quiz built in Streamlit that presented recommendations in a visual grid with confidence indicators and captured like, dislike, and selection feedback along with a simulated purchase flow. That frontend was containerized and served on Amazon ECS Fargate behind an Application Load Balancer, with images stored in Amazon ECR. Amazon Bedrock provided supplementary validation of the model’s recommendations, while Amazon CloudWatch and AWS IAM handled monitoring and secure, role-based access.
Critically, the system enforced Madison Reed’s expert business rules directly in the recommendation logic, applying hard constraints such as permissible shade-level ranges, so that the model’s learned intelligence never overrode the guardrails the brand’s colorists depend on.
The selected Hybrid DeepFM model delivered strong, business-relevant accuracy on a held-out test set of more than 12,000 customers while respecting Madison Reed’s expert constraints, demonstrating that a learned model could match the brand’s quality bar without manual rule-keeping. Beyond accuracy, the solution proved efficient to train and inexpensive to operate, and it left Madison Reed with a documented, retrainable system its own team can adapt as the catalog grows. Customer sentiment throughout the engagement was positive.
Let’s explore your high-impact AI opportunities together in a complimentary session
