loading

Business Impact

  • Improvement in item selection ratio

  • Highly agile and scalable

Customer Key Facts

  • Country : USA
  • Industry : Retail

Problem Context

The client is a provider of subscription-based food delivery services and a customer-first platform that is headquartered in New York, United States. They wanted to develop a Contextual Bandit Recommendation Engine in the Daily Harvest GCP infrastructure to suggest the best recommendation model from some given contextual features about the user and the available N recommendation models. This would increase the chance of the user adding to the cart at least one of the recommendations by the selected recommendation model (and potentially minimize the chances of an add-to-cart that leads to pause and cancellation of subscription).

Challenges

  • Efficiently evaluate new model performance with the deployed models.
  • Rigorous A/B testing that increased time to market.
Challenges

Technologies Used

Google Cloud Storage
Google BigQuery

Google VertexAI

Google Cloud Build
Google StackDriver
Cloud PubSub

Terraform
DBT
Github

Developed a Contextual Bandit Recommendation Engine in the Daily Harvest GCP infrastructure

Solution

Quantiphi developed the contextual bandit solution by using epsilon greedy with deep learning neural network and the MLOps pipeline Model. Furthermore, model training was done using Vertex AI and epsilon greedy logic was implemented in FastAPI code.

Results

  • Quantiphi helped the Daily Harvest team to develop a contextual bandit recommendation engine that suggests a recommendation model from a pool of “N” models given some user context, which maximizes the probability of a user adding a recommended item to their cart.
  • The framework is developed on GCP and fully automated (CI-CD framework to trigger the Vertex AI continuous training pipeline and a second pipeline that deploys the FastAPI) to ease the process flow.

Looking for similar project?

Let's Talk

Get your digital transformation started

Let's Talk