Chain Yourself Up To Gain Freedom For Your Enterprise!

A comment recently from my friend, Jim Stevenson , got me thinking about modern architectural patterns incorporating SLMs and microservices.

Maybe the next big thing chaining Specialised Language Models (SLMs). It's almost like AI’s version of the microservices or serverless architecture—only more flexible, more powerful, and just maybe, more fun.

And while the focus below is on chaining SLMs, having a robust architecture and orchestration model will allow you the freedom to take control of legacy and modern systems alike. I've also not covered sustainability, but it's worth noting that SLMs run at massively smaller energy costs than broader solutions like GenAI.

But what does this really mean for developers and businesses alike? .

Is Chaining SLMs Just the New Microservice Architecture?

To start, let's put things into perspective. Microservices were a breath of fresh air in a world dominated by monolithic architectures. Instead of one giant block of code, we broke down applications into small, manageable, and independent services. These microservices could be developed, deployed, and scaled independently, reducing the dreaded “single point of failure.”

Chaining SLMs is like that—but on AI steroids. Imagine each microservice not just being a static piece of code but a dynamic, learning, and adaptable entity. When we talk about chaining SLMs, we’re essentially linking together multiple AI models, each with a specialised focus, to perform complex tasks. It’s like having an orchestra where each instrument knows its part and plays it perfectly, yet is smart enough to improvise if the sheet music changes mid-performance.

So, is it just the new microservice architecture? Yes and no. It's an evolution—a step forward. While microservices were about dividing and conquering tasks, chaining SLMs is about collaboration and adaptability. It’s the difference between assembling a jigsaw puzzle and creating an evolving masterpiece that paints itself as you go along.

Orchestrating for Reliability: Keeping the AI Band Together

Now, here’s the catch. As with microservices, orchestration becomes crucial. If one SLM decides to go rogue or misinterpret its input, the whole chain can crumble. Reliability is everything.

To tackle this, we can look to orchestration tools that have served us well in the microservices world—think Kubernetes for microservices or Lambda for serverless functions. But in the realm of SLMs, we need something more nuanced. These AIs need to not only talk to each other but understand each other.

Enter advanced orchestration platforms that can manage AI workflows, ensuring each model in the chain performs its job correctly and communicates effectively with the next. These tools might include AI pipelines managed by tools like Kubeflow, or frameworks that integrate and monitor AI models like MLflow. These aren’t just deployment tools—they’re the conductors ensuring every AI in the chain plays in harmony.

Driving Business Value with Domain-Specific Models

General AI models are like Swiss Army knives—they’re good at a lot of things, but not perfect at anything. By using domain-specific models, we can drill down into the nitty-gritty details of business needs, whether that’s customer service, financial forecasting, or content creation.

These specialised models add genuine business value by being finely tuned to specific tasks. Imagine an AI model trained exclusively on legal documents—its understanding of legal language and context would far outstrip a general-purpose model. Now chain that with another model specialised in financial data, and you’re getting business insights that are precise and actionable.

This specialisation is where the real business value lies—solving problems more effectively than ever before. Domain-specific SLMs are like hiring a team of experts rather than a group of generalists. They’re not just better at their jobs; they’re indispensable.

The Issue of Unstructured Data: Do We Finally Have a Solution?

One of the biggest headaches in AI and data science has always been unstructured data. Getting it right is a herculean task, and traditional approaches often fell short. But chaining SLMs could offer a new hope.

By chaining models that specialize in different types of data—text, images, videos, and more—we can create a pipeline that processes and understands unstructured data more effectively. For instance, one model could be dedicated to cleaning and normalizing the data, another to understanding its context, and yet another to generating insights from it. The result? A much more streamlined process that gets you closer to that elusive “right” interpretation of your unstructured data.

No more endlessly wrangling your data to fit into a one-size-fits-all model. Instead, let each SLM handle what it knows best, handing off to the next model when needed, all while maintaining the integrity and accuracy of your data.

Blended Orchestration: The Best of Both Worlds?

Here’s a thought—why not take a hybrid approach? Use microservices or traditional logic where the answer is clear-cut and deterministic, and fall back on AI, especially SLMs, where the situation is fuzzy or open-ended.

For example, in a customer service application, a microservice could handle simple FAQ queries (like “What’s your refund policy?”) while an SLM steps in for more nuanced interactions (“I’m unhappy with my purchase, what can you do for me?”). This blended orchestration ensures efficiency where possible and flexibility where necessary.

Tools of the Trade: What’s Available Right Now?

You’re probably wondering what tools are already out there to help with all this. Fortunately, the tech landscape is rich with options:

1. Kubeflow: An open-source machine learning toolkit that lets you orchestrate complex AI pipelines, managing everything from data preparation to model deployment.

2. MLflow: This tool simplifies tracking, versioning, and deploying machine learning models, ensuring consistency across your AI workflows.

3. AWS Step Functions: A serverless orchestration service that can be leveraged to chain AI models in a reliable, scalable manner.

4. Apache Airflow: While traditionally used for ETL and data processing, Airflow’s extensible nature makes it a good candidate for managing AI pipelines too.

5. Ray: An emerging framework designed for building and running distributed applications, including those that involve AI models.

Each of these tools brings its own strengths to the table, allowing you to build, manage, and deploy chained SLMs in a way that suits your specific needs.

Conclusion: A Bright, AI-Powered Future

So, is chaining SLMs the new microservice architecture? In a way, yes—but it’s also something more. It’s a natural evolution towards smarter, more adaptable, and more specialised AI systems. By effectively orchestrating these models, leveraging domain-specific insights, and blending traditional microservices with AI-driven flexibility, we’re not just solving today’s problems—we’re setting the stage for a future where AI solutions are as flexible as they are powerful.

And that, my friends, is a future worth getting excited about.