This post originally appeared in Forbes.

Scaling an AI startup involves navigating complex challenges around machine learning model training costs, scaling the product to quickly onboard new customers, and maintaining high quality and reliability during version updates and code fixes.

In addition to their core ML competence, AI startup founders must eventually provide an effective ML Operations (MLOps) infrastructure for their business model to scale and thrive in a competitive AI landscape.

Here’s how to successfully tackle the latter challenge, with practical advice and the critical role of Machine Learning DevOps (MLOps) professionals during the journey.

Steps Towards A More Effective ML Ops Practice

Managing GPU cloud costs is possibly the most significant hurdle for AI startups. High and poorly predicted expenses on cloud resources can strain budgets, limit experimentation, and, therefore, accelerate burn rates. Implementing effective MLOps management strategies is essential to maintaining financial viability. There are a few areas where technical C-level AI founders can make a difference:

• Autoscaling: The cloud infrastructure that ML models use can be set up with dynamic resource allocation to adjust the number of running instances based on the current workload. Tools like Kubernetes’ Horizontal Pod Autoscaler (HPA) or cloud-specific auto-scaling services (AWS Auto Scaling, Azure VM Scale Sets) help optimize cost efficiency by scaling resources up or down as needed.

• Spot Instances and Preemptible Virtual Machines: The infrastructure can be further optimized by leveraging lower-cost, non-critical resources like AWS Spot Instances or Google Cloud Preemptible VMs for tasks like model training and batch processing. This approach can significantly reduce GPU expenses.

• Reserved Instances and Committed Use Contracts: For predictable workloads, technology leaders can consider reserved instances (AWS) or committed use contracts (Google Cloud, Azure) to receive substantial discounts compared to on-demand pricing. This requires upfront planning but offers long-term cost savings.

• Resource Tagging and Allocation: For better cost intelligence, startup leaders should consider implementing tagging for resources to track and allocate costs accurately by project and by customer. This practice can help identify high-cost areas and improve budget management and accountability.

• Seamless Data Pipeline Integration: For consistently high quality, reliability, seamless user experiences, and efficient resource utilization, implementing DataOps practices and automated data pipelines with tools like Apache Airflow and Kubeflow Pipelines can help ensure consistent data preprocessing and validation, reducing manual errors.

The Role And Criteria For MLOps Professionals

If most of the above sounds way too technical, the appropriate approach would be to build Machine Learning DevOps (MLOps) capability or hire a service provider to take care of the infrastructure setup and management.

Here are some of the key experience criteria you should test when hiring or vetting MLOps professionals and service providers:

• Deployment automation to automate the CI/CD pipeline, ensuring seamless integration, testing and deployment of ML models. This reduces manual intervention and accelerates development cycles.

• Performance monitoring skills to set up robust monitoring and logging systems and track model performance and behavior in real time, enabling prompt identification and resolution of issues.

• Continuous improvement to implement continuous retraining and model monitoring for detecting and addressing performance drift, ensuring sustained model accuracy and reliability.

• Technical expertise that includes proficiency in relevant tools and technologies (e.g., Kubernetes, Docker, CI/CD tools and cloud platforms).

Of course, proven ability to troubleshoot and resolve issues promptly, experience in optimizing resource allocation and managing costs effectively, familiarity with different types of ML models and their deployment requirements in a variety of production environments can be considered great bonus points.

ML Ops Takeaways

It is no secret that AI startup founders face many pressures and risks. As critical as they are to growing and scaling, Operations (or ML Ops) should not be a core concern. By implementing existing best-practice ML Ops strategies and leveraging the expertise of MLOps professionals, AI startup founders can effectively manage a surprising number of typical issues like costs and scaling, as well as offer high quality and reliability, usually typical at bigger enterprises.

Embracing ML Ops solutions can help AI startup founders navigate the complexities of growing their AI startup and achieve sustainable growth and success.