50%
reduction in time spent waiting for data queries to run.
22%
reduction in overall architecture costs.
The Challenge
Crisp faced two key challenges in their buying process:
- Waste & Out-of-Stock: difficulty in creating accurate demand forecasts due to hypergrowth, seasonality, and product launches.
- MLOps Maturity: the existing data platform (GCP, DBT, Airflow) lacked a mature MLOps setup, leading to over-reliance on supply chain experts and hindering model development.
Scalability was blocked by two key bottlenecks:
- Data Ingestion: The legacy data warehouse couldn’t handle the volume, leading to query times reduced by 50%.
- MLOps Ownership: Data Scientists were heavily dependent on the back-end team for model deployment, resulting in a time-to-production measured in months.
Our Approach
Following a design sprint, we designed and implemented a best-of-breed MLOps architecture focused on autonomy and scalability:
- Data Stack Shift: Migrated the data warehouse to Snowflake to manage rising volumes and improve query speed.
- Automation & Governance: Implemented Managed Airflow and dbt to streamline data ingestion and transformation with better control and ownership.
- MLOps Ownership: Introduced Deeploy for model serving and monitoring, empowering the Data Science team to deploy and iterate on models independently of the product team.
"Enjins really brought in their expertise on how to tackle the forecasting challenge. Combining general forecasting challenges with Crisp specific ones to come up with a next level forecasting design."
Eric Klaassen
Co-founder & CPO at Crisp
