To harness the power of AI and machine learning, organisations must focus on crucial aspects of model selection, optimisation, monitoring, scaling and defining success metrics.
Successfully integrating and managing artificial intelligence and machine learning into business operations has become a critical priority for organisations striving to remain l competitive in an ever-evolving landscape. However, for many organisations, harnessing the power of AI/ML in a meaningful way is still an unfulfilled dream. To address this challenge, I’ve examined some of the latest trends in MLops and compiled actionable strategies that can help solve common fML engineering challenges.
As you might expect, generative AI models differ significantly from traditional machine learning models in their development, deployment, and operational requirements. I’ll walk through these differences, which range from training and the delivery pipeline to monitoring, scaling, and measuring model success and leave you with a few key questions organisations should address to guide their AI/ML strategy.
Ultimately, by focusing on solutions, not just models, and by aligning MLops with IT and DevOps systems, organisations can unlock the full potential of their AI initiatives and drive measurable business impacts.
The core principles of MLops
Like many things in life, in order to successfully integrate and manage AI and ML into business operations, organisations first need to have a clear understanding of the foundations. The first fundamental of MLops today is understanding the differences between generative AI models and traditional ML models.
Cost is another major differentiator. The calculations of generative AI models are more complex, resulting in higher latency, demand for more computer power, and higher operational expenses. Traditional models, on the other hand, often utilise pre-trained architectures or lightweight training processes, making them more affordable for many organisations. In fact, 14% of UK organisations are still avoiding the use of ML models altogether, with cost and complexity likely playing a key role in that hesitation. When determining whether to utilise a generative AI model versus a standard model, organisations must evaluate these criteria and how they apply to their individual use cases.
Optimising models and effective monitoring
Optimising models for specific use cases is crucial. For traditional ML, fine-tuning pre-trained models or training from scratch are common strategies. GenAI introduces additional options, such as retrieval-augmented generation (RAG), allowing private data to provide context and ultimately improve model outputs. Choosing between general-purpose and task-specific models also plays a critical role. Do you really need a general-purpose model, or can you use a smaller model trained for your specific use case? General-purpose models are versatile but often less efficient than smaller, specialised models built for specific tasks.
Model monitoring also requires distinctly different approaches for generative AI and traditional models. Traditional models rely on well-defined metrics like accuracy, precision, and an F1 score, which are straightforward to evaluate. In contrast, generative AI models often involve metrics that are a bit more subjective, such as user engagement or relevance. Good metrics for genAI models are still lacking and it really comes down to the individual use case. Yet nearly half (47%) of UK organisations are skipping runtime security scans, highlighting a critical gap in how models are being monitored in production. Assessing a model is very complicated and can sometimes require additional support from business metrics to understand if the model is acting according to plan. In any scenario, businesses must design architectures that can be measured to make sure they deliver the desired output.
Evolving ML engineering practices
Traditional machine learning has long relied on open source solutions, from open source architectures like LSTM (long short-term memory) and YOLO (you only look once), to open source libraries like XGBoost and Scikit-learn. These solutions have become the standards for most challenges thanks to being accessible and versatile. Interestingly, the UK shows greater caution here, 38% of organisations restrict public software downloads, 10% above the global average, which may impact their ability to fully leverage these open source tools. For genAI, however, commercial solutions like OpenAI’s GPT models and Google’s Gemini currently dominate due to high costs and intricate training complexities. Building these models from scratch means massive data requirements, intricate training, and significant costs.
Despite the popularity of commercial generative AI models, open-source alternatives are gaining traction. Models like Llama and Stable Diffusion are closing the performance gap, offering cost-effective solutions for organisations willing to fine-tune or train them using their specific data. However, open-source models can present licensing restrictions and integration challenges to ensuring ongoing compliance and efficiency.
Scaling ML systems with efficiency
As more and more companies decide to invest in AI, there are best practices for data management and classification and architectural approaches that should be considered for scaling ML systems and ensuring high performance.
Utilising internal data with RAG
Important questions revolve around data: What is my internal data? How can I use it? Can I train based on this data with the correct structure? One powerful strategy for scaling ML systems with genAI is retrieval-augmented generation. RAG is the ability to use internal data to change the context of a general purpose model. By embedding and querying internal data, organisations can provide context-specific answers and improve the relevance of genAI outputs. For instance, uploading product documentation to a vector database allows a model to deliver precise, context-aware responses to user queries.
Architectural considerations for scalability
Creating scalable and efficient MLops architectures requires careful attention to components like embeddings, prompts, and vector stores. Fine-tuning models for specific languages, geographies, or use cases ensures tailored performance. An MLops architecture that supports fine-tuning is more complicated and organisations should prioritise A/B testing across various building blocks to optimise outcomes and refine their solutions.
Defining success with metrics
Aligning model outcomes with business objectives is essential. Metrics like customer satisfaction and click-through rates can measure real-world impact, helping organisations understand whether their models are delivering meaningful results. Human feedback is essential for evaluating generative models and remains the best practice. Human-in-the-loop systems help fine-tune metrics, check performance, and ensure models meet business goals.
In some cases, advanced generative AI tools can assist or replace human reviewers, making the process faster and more efficient. By closing the feedback loop and connecting predictions to user actions, there is opportunity for continuous improvement and more reliable performance.
Solutions over models: A holistic approach
The success of MLops hinges on building holistic solutions rather than isolated models. Solution architectures should combine a variety of ML approaches, including rule-based systems, embeddings, traditional models, and generative AI, to create robust and adaptable frameworks.
Organisations should ask themselves a few key questions to guide their AI/ML strategies:
● Do we need a general-purpose solution or a specialised model?
● How will we measure success and which metrics align with our goals?
● What are the trade-offs between commercial and open-source solutions, and how do licensing and integration affect our choices?
Ultimately, the takeaway is clear: AI and ML are no longer about developing standalone models, they are about building holistic, interconnected solutions.. These solutions comprise intricate architectures where each component influences the overall user experience and success metrics derived from them. As MLops continues to evolve, organisations must focus on crafting scalable, metrics-driven systems. By leveraging the right combination of tools and strategies, businesses can unlock the full potential of AI and machine learning to drive innovation and deliver measurable business results.
About the Author
Yuval Fernbach is VP, CTO MLops at JFrog. Deliver Trusted Software with Speed. The only software supply chain platform to give you end-to-end visibility, security, and control for automating the delivery of trusted releases. The massively scalable, hybrid JFrog Platform is open, flexible, and integrated with all the package technologies and tools comprising the software supply chain. Organizations benefit from full traceability to any type of release and deployment environment including ML models, software that runs on the edge, and software deployed in production data centers.
