AiSolve India

How do you architect a product after getting ideas? How do you develop an architecture for an AI product that ensures reliability, explainability, and trustworthiness? How do you go beyond applying LLMs in an AI product?

In the previous lesson, we learned the steps to initiate good product development that delivers value to customers. One critical step is product architecture. In this lesson, I will describe the methods I follow to architect a product.

What is the difference among framework, architecture and process in a product? Let me explain it with an example. A framework guides you; it tells you what is needed. But it does not tell you how to do it. An architecture defines the structure; it tells you what it has and how the parts are interconnected. A process refers to execution: how actions are carried out to deliver an output.

To start, I define the architecture. While defining the architecture, I ensure that the following artefacts are well defined.

1. Tools: Collection of tools, databases, platforms.
2. Clusters: The architecture contains multiple clusters. Each cluster contains similar entities. For example, all platforms the product interacts with can be clustered together, and all AI Agents can be clustered as one, etc.
3. Connections: All tools and clusters are connected through directed acyclic graphs (DAGs).

The architecture gives an overview of what clusters the product should have and how they interact with one another, including the tools. It also shows all the product’s functionalities.

After designing the product architecture, it is time to break it down further. In this stage, I create processes. A single product can have many processes. While the product architecture describes the functionalities, processes describe the features. To create a process, I use the concept called VSM (Value Stream Mapping). VSM give you a detailed outline of the process, current state and future state. The current state tells where we stand, and the future state tells what we should do next to solve a pain point. To create a process, I ensure the following artefacts are well defined while defining the VSM.

1. Steps: I write down the detailed steps and connect them with graphs. I aim to optimise the process to have the fewest loops, ideally none. Loops can make the steps unstable; however, when a product contains multiple clusters of processes that interact, automatic feedbackautomatic feedback loops form. However, designing with the minimum number of loops is a best practice I consistently follow.
2. Current State: For each step, what we have currently. This reduces the chances of duplicate work and reinventing the same wheel. Often, we can reuse existing solutions with a slight reconfiguration. 
3. Future or Next State: For each current state, I identify what is missing or what the pain is. To solve it, I work with the team to figure out what’s next. That defines the future or next state. 
4. For each future state, I write what value it delivers. For example, reduced cycle time, reduced defect leakage, what this new feature solves, etc.
5. Tools: Finally, I consider tools connected to each step, with a description of what they do.

With this, I have a detailed architecture of the product and processes that fit the product. This forms the starting point of product prototyping, as mentioned in the previous lessons.

As an AI generalist, one must know the above artefacts fluently. Processes vary by industry, but the artefacts mentioned above remain essential for designing a good product architecture. Readers can pause here and do some simple research on the VSM (Value Stream Mapping). In the next chapter, we will learn how to add AI functionality to the architecture and processes above to solve complicated problems reliably, rather than just calling an LLM in the steps to solve simple problems.