The Challenge
The ambition to scale from local diver-operated campaigns to thousands of kilometers of drone-operated coastline created immediate structural blockers. PlanBlue’s existing infrastructure was built for a startup context — robust enough for early-stage operations, but not designed to absorb the exponential growth in data volume, team size, and processing complexity that lay ahead. Three challenges stood in the way:
- Pipeline scalability at a new order of magnitude: The existing processing pipeline was not architected to handle the surge in hyperspectral imaging data generated by autonomous underwater vehicles. Without significant re-engineering, the system would become a bottleneck at exactly the moment PlanBlue needed to accelerate.
- Manual bottlenecks slowing operations: Critical steps in the data processing pipeline relied on manual triggers, introducing unnecessary delays and increasing the risk of human error. As campaign volume increased, this approach would become operationally unsustainable.
- MLOps maturity lagging behind team growth: A growing team of data and ML engineers lacked a standardized development environment, making collaboration increasingly difficult. Without proper model tracking, version control, and environment isolation, maintaining and improving algorithms at scale was becoming a real friction point.
Our Approach
Our Approach
Following a four-week design sprint to validate existing algorithms and blueprint the infrastructure required to support PlanBlue’s hardware expansion, Enjins partnered with PlanBlue to transition them from a startup setup to a production-grade data organization.
Design Sprint as Diagnostic Foundation
We began by assessing the performance of PlanBlue’s existing computer vision models across diverse water and lighting conditions. The models proved robust — confirming that the path to scale ran through infrastructure, not algorithmic redesign. The design sprint produced a concrete infrastructure blueprint and implementation roadmap that guided the entire development phase.
Scalable Data Platform and Pipeline Re-architecture
We re-architected the processing pipeline for the demands of autonomous drone campaigns. By wrapping PlanBlue’s existing Kedro modules in AWS Batch, we enabled massive parallelization of data processing steps — allowing the system to scale compute automatically in the cloud as data volumes grow. Storage and metadata management were standardized across S3 and MongoDB, ensuring consistent and efficient data retrieval across the entire pipeline.
Event-Driven Orchestration with Airflow
To eliminate reliance on manual triggers, we introduced Airflow for flexible, event-driven orchestration of processing steps. Critically, the implementation preserved PlanBlue’s need to manually intervene or adjust steps based on data-specific conditions or client preferences — giving the team autonomy without sacrificing automation.
MLOps Professionalization
To empower the growing internal team, we implemented a mature MLOps stack. Strict isolation between development and production environments was established, enabling engineers to build and test improved models without risking disruptions to live pipelines. MLFlow was introduced as a central platform for model tracking and registration, replacing a fragmented process of gathering results across sources and conversations. Terraform was used throughout to ensure the infrastructure is fully reproducible, maintainable, and easy to extend as the team scales.
Team Enablement for Long-Term Independence
All infrastructure decisions were made with the internal team in mind — the people who would work with and maintain these systems day to day. By the end of the engagement, PlanBlue’s in-house AI team was equipped to continue scaling the system independently, ensuring the collaboration had lasting impact beyond the project itself.
