## Overview This MLOps Community podcast episode features a discussion between Amita Arun Babu Meyer, an ML Platform Leader at Klaviyo, and Abik (a senior managing consultant at IBM), moderated by Demetrios. The conversation provides a comprehensive overview of MLOps maturity levels within businesses and how organizations can tie technical capabilities back to measurable business value. The discussion covers both traditional ML and emerging LLM operations, offering perspectives from both technical and product management viewpoints. ## MLOps Maturity Levels Framework The speakers outline a three-tier maturity model for MLOps that has become increasingly relevant as organizations move beyond experimentation to production-grade machine learning systems. ### Level Zero: Manual and Ad-Hoc At the foundational level, organizations are just beginning their MLOps journey. The characteristics of this stage include: - Data scientists orchestrate experiments manually, covering data preparation, model training, evaluation, and validation in isolated development environments - Deployment is performed manually with minimal or no versioning - The entire ML lifecycle can span several months - There is limited standardization across teams - Code repository usage and versioning practices are inconsistent The speakers note that even mature IT organizations with sophisticated infrastructure often find themselves between Level Zero and Level One, highlighting how challenging this progression can be. ### Level One: Semi-Automated with Standardization The transition from Level Zero to Level One involves introducing standardization and partial automation. Key improvements include: - Standardized templates for experimentation and development, such as Docker images that can be used for both experimentation pipelines and deployment - Alignment between data engineering and data science teams using common frameworks - Automated orchestration of the experimentation pipeline (data prep, model training, evaluation) - Packaging and deployment of entire pipelines rather than just code—a crucial distinction from traditional DevOps - Introduction of model registries for versioning, which provides team synchronization and the flexibility to roll back to previous model versions when business requirements change Amita emphasizes that standardization reduces time for developers significantly, increases collaboration, and decreases back-and-forth communication overhead between teams. ### Level Two: Full Automation with Continuous Monitoring The aspirational state involves complete automation of both integration and serving: - Automated orchestration using containerization technologies (Docker, Kubernetes) - Model registry integration for comprehensive versioning - Continuous monitoring of both data and models in production - Automated retraining triggered either by schedule or by data/performance changes - Detection of data drift, skewness changes, and other anomalies that trigger model updates - The ability to automatically adjust hyperparameters based on performance degradation Abik notes that achieving Level Two remains elusive for most organizations. Even clients he has worked with since 2018 are typically somewhere between Level One and Level Two after several years of development. ## LLM-Specific Considerations The discussion addresses how the emergence of LLMs changes the MLOps landscape, identifying several unique challenges: ### Data Pipeline Complexity LLMs require massive datasets, and the speakers note that having robust data pipelines for accessing both internal and external data sources becomes even more critical. The progression from Level Zero to Level One for LLM deployments involves: - Establishing data catalogs that provide insights into available data sources - Understanding data lineage and schema information - Moving from one-time batch data acquisition to continuous, refreshed data access ### GPU Cost Concerns One significant barrier to full automation for LLM deployments is the cost of GPU compute. Organizations using models like GPT-4 for experimentation often remain cautious about moving to complete automation because GPU costs can "snowball" during high-usage periods. This economic consideration represents a practical constraint that doesn't apply as strongly to traditional ML deployments. ### Compliance and Privacy The speakers highlight compliance as a critical consideration, particularly when using third-party LLM APIs. Organizations must perform due diligence to ensure that external model providers handle data in accordance with company policies and regional regulations. This becomes especially complex in global organizations where different countries have different privacy laws—the example is given of Canadian healthcare data that cannot leave Canada, creating challenges for building RAG chatbots or other LLM applications that might inadvertently access protected data. ## Tying Technical Improvements to Business Value Amita provides the product management perspective on translating MLOps improvements into business metrics: ### North Star Metric The ultimate goal of any ML platform is to reduce the time from ideation to production for data scientists. This velocity metric serves as the primary measure of platform success. ### Design Thinking Approach The speakers advocate for working backwards from customer needs, which in the ML platform context means: - Identifying primary customers (data scientists), secondary customers (ML engineers), and tertiary customers (BI engineers/analytics roles) - Understanding pain points across each phase of the ML lifecycle - Mapping improvements to business impact by quantifying the cost of unresolved issues ### Translating Time to Dollars A practical example is given for quantifying the business impact of technical decisions. If data scientists need to learn Spark (because they're comfortable with Python but the organization uses Spark for large-scale distributed compute), the calculation might be: - Number of data scientists who need the new skill × hours required for training × hourly cost = direct training investment - Plus the hidden costs of scientists making mistakes while learning, time diverted from innovation, and the risk of errors in production This framework allows technical leaders to communicate the value of platform investments in terms that resonate with business leadership. ## Persona-Specific Learning Paths The discussion outlines what professionals from different backgrounds need to learn when entering MLOps: ### Data Engineers Entering MLOps Data engineers already understand data acquisition, transformation, and EDA. They need to develop: - Understanding of various ML model types - Knowledge of what format and structure data scientists require - Awareness of how data preparation impacts model performance ### Data Scientists Expanding into Operations Data scientists excel at experimentation and development but need to learn: - Parameterization practices for production code - Code repository and versioning workflows - YAML configuration (moving beyond Jupyter notebooks) - Understanding of deployment constraints ### DevOps Engineers Moving to MLOps DevOps professionals understand CI/CD but need to grasp: - The iterative nature of ML pipelines - Hyperparameter tuning concepts - How changes in data translate to necessary model adjustments - The differences between deploying code versus deploying entire ML pipelines ### Product Managers Transitioning to ML Products Amita shares her own experience, noting steep learning curves in: - Data science terminology and concepts - Understanding the difference between real-time and batch pipelines at a code level - Model architecture decisions (e.g., why some models don't need ground truth data) - Building relationships with data scientists by asking informed questions ## Model Monitoring Evolution Across Maturity Levels The conversation addresses how monitoring practices evolve: - **Level Zero**: Monitoring is rudimentary, with long feedback cycles. Business users typically flag performance issues, triggering manual investigation of data quality or model drift. - **Level One**: Basic monitoring infrastructure in place, but ground truth data pipelines may be incomplete. Teams often resort to "scrappy" solutions like manual sampling of model outputs. - **Level Two**: Dedicated monitoring tools track multiple KPIs continuously. Automated alerts trigger retraining when performance degrades. Cloud platforms provide increasingly sophisticated monitoring capabilities. For LLMs, the speakers note that traditional accuracy metrics don't translate well. Evaluating summarization quality or other generative outputs requires different approaches, explaining the proliferation of LLM evaluation tools in the ecosystem. ## Key Takeaways The discussion concludes with a critical insight: ML and AI teams are often viewed as cost centers rather than profit centers by leadership. The responsibility falls on technical practitioners to clearly define and communicate their business impact. Whether through velocity metrics, cost savings calculations, or revenue attribution, the ability to tie MLOps improvements to business value is essential for securing continued investment in ML infrastructure and advancing through maturity levels.