Yet Another Generative-AI blog…?
Why on earth start another Generative AI blog?
Since ChatGPT burst onto the scene in November 2022, I’m sure your social feeds, much like mine, have been filled with content predominantly about generative AI—and for good reason. Over the last few years, this emerging technology has been adopted on a scale few could have predicted, leading to an urgent rush to understand, contextualise, and, above all, productise solutions led by large language models. This has spurred the creation of numerous new YouTube channels, blogs, and LinkedIn posts to cater to this still-nascent technology.
My day job as a Microsoft Cloud Solution Architect revolves primarily around application development and process automation. I am by no means an AI specialist. However, since March of last year, I have had the fortune to architect and build several real-life applications infused with LLMs, which has been both immensely enjoyable and challenging. I believe that in terms of building solutions atop LLMs, we are all very much learning and will continue to do so for some time. While I will aim to explore new and emerging features, I also think it’s important to discuss how we can turn these exciting features into applications that deliver real benefits. So, as I continue to develop real-life applications alongside various proofs-of-concept, I want to document my learning journey as a means to retain what I’m learning, and also in the hope that others dipping their toes into this area will find the content beneficial.
What I want to focus on…
Custom LLM applications
Example custom LLM application leveraging Azure OpenAI Assistants
When I discuss Generative-AI business applications with customers and other professionals in the Microsoft Business Applications sector, I am frequently asked why they should consider custom LLM applications given the capabilities of off-the-shelf solutions like Microsoft's Copilot ecosystem. Indeed, for many use cases, the off-the-shelf products are perfectly adequate. However, by opting for these, we as users and developers sacrifice a significant degree of interaction with the underlying models and orchestration of their responses. This often results in content that is somewhat generic and not fully applicable to the specific domain in which we are operating.
This issue becomes critical for organisations that interact with external audiences where maintaining a consistent tone of voice and brand continuity can significantly impact the company’s reputation and financial performance. For instance, consider a chatbot integrated into the website of a high-end boutique fashion retailer based in the UK. It is reasonable to expect that the chatbot should communicate in a manner consistent with the retailer’s desired tone of voice. Achieving this level of tailored communication is challenging with off-the-shelf products, where developers rarely have access to the system prompts, let alone the ability to fine-tune and control the orchestration of responses. This is one of the many reasons why I feel we need to better bridge the gap from exciting proofs-of-concept LLM applications to robust production ready products with custom solutions.
The Azure Generative-AI ecosystem (especially Azure OpenAI)
Overview of Microsoft led Generative AI solutions
My most extensive implementation experience lies within the Azure AI stack, encompassing everything from custom solutions developed on Azure OpenAI to simpler, low-code Copilot applications. Microsoft often champions a low-code-first approach, which indeed suits many customers and a variety of use cases. However, these solutions have inherent limitations due to restricted access to the underlying system prompts and their orchestration, making the need to build more tailored and flexible solutions on top of enterprise grade platforms such as Azure OpenAI.
Whilst I often work with Microsoft’s lowcode tools, there are numerous prominent voices in the Microsoft Business Applications community that extensively cover Copilot Studio. Yet, when it comes to developing applications using Azure AI services such as Azure OpenAI and Azure AI Studio, there seems to be a notable dearth of detailed content.
I am eager to share my experiences as I move beyond proofs-of-concept to create production-ready, robust applications. Additionally, I look forward to exploring new features as they are released and integrating them into practical, impactful solutions.
New Generative-AI tech that I find intersting
Generative-AI beyond Microsoft
From the lightning speed inference of Groq to intelligent orchestration tools such as LangChain, there’s a lot to explore beyond the Microsoft eco-system.
So what’s next…?
Over the past few weeks, I have been working on developing an application that leverages the OpenAI Assistants API within a full-stack environment. A sneak preview of this project is available here. I will soon share more details on how you can build your own version of this application and discuss whether the Assistants API is the right framework to invest your time in for future development endeavors.
Additionally, I've been exploring various Generative AI platforms, LLMs, and libraries, including Groq, Llama-3, and LangChain. These explorations have sparked numerous content ideas, which I'm excited to share with you.
In the meantime, thanks for stopping by and reading my first blog post. Stay tuned for more updates in the next one!