Two thought processes intersected in my mind this morning. I started the day refreshing a streaming system concept for my weekly “Edge AI Bytes” video. After that, I worked on a hands-on project about leveraging a LLM model on your interactive website. As I was sipping tea with my wife, these two aspects intermingled to bring forth the idea that architecting streaming systems, or for that matter, solution architecture, can be almost automated with Generative AI. In fact, Generative AI can do a much better job than most human solution architects can do.

The article will be short since the basic premise of my postulation is simple. A solution architect leverages few critical steps of expert knowledge to formulate an architecture:

• 1: An understanding of the current state of infrastructure
• 2: Documentation of required future state
• 3: Tapping expert knowledge of architectures needed in similar scenarios, leveraging 1 and 2
• 4: Map the needed architecture with available components (like services by a hyperscaler) to design the required architecture
• 5: Build the final architecture

As you can imagine, Steps 1 and 2 are structured information easily absorbed by a LLM model. The same applies for step 4. A structured and defined list of service components must be used to design an architecture, with precise functionality defined for each.

The critical training piece here is the expert knowledge. That is where the training needs to happen. A vast dataset of data points from 1, 2, 3, and 4 needs to be leveraged to train a deep-learning model. At a high level, in simple terms, 4 is the target, whereas 1,2, and 3 are the inputs. If a vast amount of data can be created, the training should not be difficult. The other positive side of this is that AI algorithms can be used to build training datasets as well.

As I promised, this will be a short article. But the postulation can be really powerful. It can help accelerate the digital transformation timelines. However, like most digital transformation initiatives, the challenge lies in building a robust underlying dataset. Standardizing information in steps 1,2 and 3 is the challenging part. But the fruits of the effort can be lovely.

Have a nice Thanksgiving !!!