After working for nearly a year on a Gen AI initiative, here’s what I’ve learned…
When it comes to productionizing GenAI initiatives with business goals, it goes beyond just coding the solution. There are factors that come into picture and need to be in place for a successful and useful productionization. Drawing from my experience, here are a few key points and execution steps to help you build a successful GenAI product.
Firstly, while choosing the right GenAI use case for your organization/business, you need to ensure that you have the data and it’s right dimensions for the GenAI use cases you wish to work with -- your Generative model is only as good as your data. This is very crucial aspect since, if your data pipelines aren’t set or if your data is not well labelled your Gen AI initiative will fail!
Once, you have identified the exact use case, by considering various factors including - cost, data, user adoption, and the benefit, now’s the time to productionize your use case to reality.
Here’s a walkthrough the productionization :
User Roadmap
Start with user roadmaps with pain points and the introduction of Gen AI will ease those and what will be the benefits. That’s how you’ll be setting the expectations but also being close to reality.
Here’s an example,
User Interface and adoption
It’s no hidden truth that you might have world’s most efficient data model running behind but if the user interface is complex and unappealing, it won’t work. To overcome this, I would suggest always start with a PoC with a bare minimum User Interface that gives you a basic functionality like input, upload, download, output etc.
- Remember, a bare minimum UI is one of the main reasons of chatgpt’s success in extensive reach and usage!
Best model choice
Decide your model based on your use case and the intended output, for example - a chatbot or content generator or summarizor or image extractor. Depending on this, you might also need a multimodal model or a llm, look carefully and choose your model and avoid unnecessary extra costing.
Scaling the solution
Once you validate your PoC, now it’s time to scale your solution — technically as well as on adoption. You might have validated your PoC on 5-20 documents and now there’s a large set of documents to be processed. Here’s where your technical development plays a major role on the software devlopment and DevOps/MLOps enhancements to the model processing.
Changing Components of the Design
This the most important part to remember while you are going through all the above steps.
The field of AI is so fast paced that by the time your PoC is finished, you may have a better component at a cheaper price, and you might have to change your architecture and development accordingly.
Then how to change the components so quickly and still keep moving forward without loosing much time ?
Loosely coupled architecture is the way to handle changing components in fast pace. This helps in decoupling any of the components and adding new ones.
Woohoo!!! Now your cost, time and effort optimized Gen AI solution is production ready!
Give a clap if you appreciate the article and share more if you find this insightful. (After all, “Sharing doubles the knowledge and not sharing makes it back to zero!”😉)
— Siddhi!