Driving More Efficient Nature Identification with the Latest in AI Technology

Fully Automated ML Pipelines as a Robust Foundation for Faster, At-Scale AI/ML

Earth.com is the premier destination for people who care about earth and want to make a difference. It is home to a vast trove of information regarding life on our planet, and it provides a functioning reference library to satiate your curiosity or assist in research projects.

As a continuation of its life cataloging initiative, Earth.com has developed EarthSnap, an AI-powered application that enables users to identify all types of plant and animal species via a mobile phone camera. EarthSnap is already available in the Appstore and Playstore.

Recognizing the strategic importance of EarthSnap for their future expansion, the leaders of Earth.com were looking to enhance the application with advanced machine learning and computer vision models, and to release the application with production-ready AI/ML as quickly as possible. To accomplish that, they collaborated with an AWS Premier Consulting partner that delivered image classification models on Amazon SageMaker.

While the accuracy of the models was satisfactory, they were delivered across various notebooks, which required manual sequential execution to process the available data and retrain the model. The deployment of endpoints also had to be done manually. Earth.com did not have an in-house ML engineering team, which meant that they would not be able to add new datasets featuring new species, improve existing models and release new ones, and scale the ML solution. The delivered solution could not be used as efficiently as intended without an end-to-end machine learning infrastructure.

Earth.com was looking for a new strategic partner to accelerate and streamline the delivery of EarthSnap to market. Provectus, an AWS Premier Consulting Partner with competencies in Machine Learning, Data & Analytics, and DevOps, was chosen for the role based on its expertise in full-scale AI/ML delivery, ML infrastructure development, and MLOps projects.

Provectus proposed to approach the EarthSnap project through a series of engagements, summarized as follows:

• Deliver fully automated, end-to-end ML pipelines to facilitate the release and deployment of new models
• Based on the pipelines, build an infrastructure for the MLOps platform, with all components required for streamlining the AI/ML work
• Introduce Managed AI Services to cover ML infrastructure provisioning, maintenance, cost monitoring and optimization
• Bring the EarthSnap application to its desired state through AI/ML work, data/DB operations improvements, and DevOps best practices

Given the time constraints, Provectus had to ensure perfect coordination of all development and business units. Thankfully, we were able to meet every goal within the designated time frame.

From Manual ML Work in Notebooks to Automated ML Pipelines, to Managed MLOps Platform

Provectus initiated a series of discovery sessions designed to figure out what had been done by Earth.com’s previous partner. We looked into the existing codebase, inventorized the notebook scripts, and set up an evaluation framework to reproduce the results of previous experiments. We proceeded by retraining the model on extended data with new classes, species, and more images. The model was packaged for deployment on the Amazon SageMaker endpoint.

At that stage, it was clear that:

EarthSnap had all necessary components for data ingestion, preprocessing and model training. However, they were available as disjointed Python scripts and notebooks, which required lots of manual heavy lifting done by engineers. In the meantime, Earth.com needed a solution for non-technical professionals.

Provectus continued to work on the AI/ML parts of the application. The models were operationalized for production usage. Two end-to-end ML pipelines were designed and built, to make it easier to gradually improve and extend the models with new species. Thanks to the pipelines’ enhancements and seamless integration with CI/CD tools, Earth.com could roll out new versions of models, much faster and easier.

Provectus fully automated the training and deployment of the models. In practice, it meant that the release of new models did not require technical support. A new model could be released with a single click from an Earth.com employee.

EarthSnap’s machine learning infrastructure was ready to train, fine-tune and deploy new models, faster and on a larger scale. We later worked on a candidate model, which resulted in success. Here ML engineering benefited from experiment and lineage tracking, existing training, evaluation, and deployment pipelines.

At that point, Earth.com had an advanced ML solution, but they were ready to move forward. Provectus offered to enhance the solution with its Managed MLOps Platform, to be delivered as part of Managed AI Services.

The Managed AI Services offering included:

• Maintenance and support of the existing ML infrastructure, with a focus on efficiency and cost optimization
• Implementation of components for access management, cost monitoring, networking, and compliance support
• Application of industry best practices for CI/CD, DevOps, and MLOps
• Introduction of data monitoring, model quality monitoring, dataset releases, and model retraining
• 24/7 troubleshooting and remediation of any issues that emerge during model usage

We prepared all documentation describing all aspects of working with AI/ML components and data, to empower the Earth.com team to run the system on their own, if needed.

Provectus was also responsible for the mobile app and backend components of EarthSnap. Best practices for CI/CD, DevOps, and application development were used to ensure the highest quality of service, and the timely full-scale delivery of the updated application to market.

Driving AI/ML Efficiencies, Optimizing Administrative Costs, and Ensuring Customer Satisfaction

Provectus met all of the ambitious goals set by Earth.com within a designated time frame. In just three months, EarthSnap received a modernized ML solution, and the mobile application was made ready for public release.

The modernized ML infrastructure enabled Earth.com to:

• Minimize engineering heavy lifting
• Reduce administrative costs project-wide
• Streamline technology processes

Provectus was able to achieve better performance, and ensured the adaptability, resilience, and security of the application.

As part of Managed AI Services, we reduced the infrastructure management overhead, established well-defined SLA and processes, ensured 24/7 coverage and support, and increased overall infrastructure stability, including production workloads and critical releases. We initiated a series of enhancements to deliver our Managed MLOps Platform and improve ML engineering.

By collaborating with Provectus, Earth.com was able to release the EarthSnap application on both the Appstore and Playstore ahead of schedule. That was a milestone achievement for the team, and for the company as a whole.

The Provectus team hopes that Earth.com will continue to modernize EarthSnap with us.

The next steps of our cooperation include:

• Finding new ways to add datasets and handle data
• Adding advanced monitoring components to pipelines
• Enhancing model retraining
• Introducing human-in-the-loop

We look forward to powering the company’s future expansion, to help them connect users with other earth-minded people from all corners of the globe.

Moving Forward

1. Learn more about our MLOps Platform, Managed AI Services, and ML Infrastructure
2. Explore more customer success stories covering MLOps & Managed AI: Appen, GoCheck Kids, VTS, Lane Health
3. Apply for Machine Learning Infrastructure Acceleration Program, to get started