What is a Software Solutions Architect?
My name is John Sebastian Lemay. I’m a software solution architect, and I’ve been doing this for 15 years. My role is to help you make more money through technology, specifically by building systems and the process to get that done. It involves understanding first – and that means we take your knowledge of your industry, business, and clients. We mix that with our understanding of technology, what’s available, and what’s feasible. We combine these two things to build a system that will take you where you need to be. Many of the systems I’ve been building recently involve AI integration with artificial intelligence, so we will be tackling that topic today.
So, I think the first step to getting started, which is important, is understanding what AI is because I believe there’s a lot of marketing out there and some claims are slightly misleading.
How Does AI Actually Work?
Everybody’s talking about ChatGPT. Let’s have a crash course on how AI works.
So, ChatGPT was born out of a process we have been using for a long time, called machine learning. Machine learning is taking data, giving it to a computer, and using an algorithm to teach the computer to recognise patterns in that data. So, that’s what we call training a model.
An example would be taking pictures and using a machine learning algorithm to get the computer to look at these pictures and distinguish the patterns that make a picture/photo of a person versus a picture/photo of something else. As you train this model, it starts remembering and recognising these patterns to a point where later you can show it a picture it’s never seen before. Because of those patterns it has learned in its artificial brain, it can tell you if it’s a photo of a person or not.
Now, these teaching algorithms and learning governments have evolved, and they’ve been able to apply them to natural language instead of pictures or other types of data. So, in the case of language models like ChatGPT, they were able to teach the computers to understand the patterns of natural language and the relationships between words and whatnot, and, at first, they were able to get these language models to complete sentences. They would give it the beginning of a sentence and they would be able to figure out what is likely to be the end of that sentence. Then, they just kept fine-tuning that to a point where we can publish a question to a language model, and it will be able to produce a human-like answer.
That’s what language models are based on – their knowledge of language patterns. They’re able to digest language natural language and produce natural language answers, which is amazing that we can do that now – but I do want to add a caveat about this, which is that these language models this is all that they do; it’s a machine that takes text and produces text. It’s true of all AI; these models take an input and can create an output now. What they do is impressive, but it is all they do, and we’ll get back to that later in the presentation; it’s just to take note of now.
Why is this revolutionary? Why is this amazing? Well, it turns out that for the longest time, computers were very strict about the kind of data they were willing to work with. I’ll give you a simple example. If you’re working with Microsoft Excel, you must work with an Excel file; there’s no other way. If you’re using software or a system needing your information, you must fill out a structured form. You will have to fill in different fields, and if you want to trigger an action, you must click a button. Everything was super specific and strict, and now that we have these language models, we can deal with natural language, which is a lot more flexible. This opens the door to many incredible things, like the ability to apply data processing to natural language. Still, it also opens the door to new ways for users to interact with systems, opening many opportunities.
How Have Companies Leveraged AI?
We will look at how some companies have largely leveraged this new capability.
For example, there’s a startup here in Hong Kong that I’m involved with called Lobahn Connect, and their mission is to connect professionals who want to build their careers with employers who are seeking good talent for their job listings. We have an algorithm that computes the compatibility between candidates and employers. This algorithm needs a lot of data to calculate a good, reliable score. What happens is we need to collect all that information from the users, so they must fill out forms. However, now that we have language models, when these people give us their CVs or get a job description, we can ask the language model to extract a lot of information from those, regardless of how it is structured. Because it’s natural language, it means suddenly, we can shave off 60% of the time it takes to onboard these users because we can extract information from these documents. We couldn’t do that before.
Another example would be Impala, a company here in Hong Kong that produces technical documentation instruction manuals. They would get products from manufacturers with accompanying reference materials, and the technical writers would have to look at all these materials to figure out what to write in the instruction manuals. Because we have language models, we can predigest all these materials from the manufacturer and prepare a draft of the instruction manual and all the information mined from these reference materials. This accelerates the work of these technical writers who just now have to revise and fill in the gaps – shaving off 80% of the time it takes to produce a user’s guide.
Lalamove, a delivery services company, have an app where users can publish written reviews about their services. For the longest time, they had to rely on staff to look at these reviews to figure out what people were happy or unhappy with. Now that we have language models, we can automate that process because the models can digest these written reviews and extract information such as what people are happy and not happy about. In this case, we’re talking about full automation of the process of making these reports based on user reviews. So, we’re talking about shaving off 95% of the time it takes to compile that data and saving costs, too.
Don’t Just Get AI And Fire All Your Employees!
AI can do amazing things, and that’s very exciting, but also, if you think this is so good, I’m just going to pay for an AI license, fire half my staff, and everything will go great – that’s not going to happen! Let me tell you what’s going to happen if you do that….
An article from Forbes in July of this year reveals that 77% of employees have reported that AI has increased workloads and hampered productivity.
So, what happened? The truth is these companies just took AI and released it in their business and said go. They didn’t know what you know today: AI alone doesn’t do much. What it does is impressive, but it doesn’t do much besides taking an input and producing an output. So, they lost out, and to understand and leverage AI, you must see AI and perceive it as a component of a bigger system. And that bigger system comprises traditional IT components assembled like we were building systems before AI became popular.
It needs to be able to connect to a database, have a UI in front, and integrate with the systems you’re already using – that is the key to success! If you want to see the gains we’ve discussed, you need to see AI as a tool that becomes integrated as part of a greater system. Knowing this lets you tackle the big question of who’s getting replaced. Are we all getting replaced now?
Will AI Replace Us?
AI, by virtue of being a new technology, will replace some people the same way that ATMs have replaced a number of bank tellers. But now that you’re also aware of the limitations and the amount of work involved in making AI work, being part of a bigger system and whatnot, it’s not going to replace everyone.
What draws the line between what it’s going to replace and what it can’t is context because, as intelligent as these language models seem, they can’t read your mind. To produce good answers, they need to know as much as possible about what they’re supposed to do. This means you need to give them many instructions to produce good answers.
Imagine you’re a manager or an executive – every decision you make involves many factors. You’ve got years of experience, you’ve got lessons learned, you know your customers. If I put you in front of a chat box and told you to write your whole life story, write everything you’ve ever learned and consider this when you make decisions, that would not be possible. In truth, AI is better suited for tasks that don’t require too much context to complete.
For example, simple data entry just asks that you have an input document and do something with it. There’s not too much context involved. But as soon as you’re writing emails or making big decisions, the context becomes too big to fit into what you can tell the AI. That’s basically what prevents AI from really taking over and replacing everyone.
Should We Be Fearful of AI?
So, those of you who joined it with fears, I hope they have been assuaged a little. AI is a tool that can assist you in your tasks. In the same way today, it would be difficult to find an office job if you don’t know how to use Microsoft Word, Excel, or even PowerPoint. I think in the world of tomorrow, if you don’t know how to leverage these tools, they might not replace you, but you might get replaced by somebody who does know how to make the best use of them. That’s the key takeaway here, and so I hope that gave you a good overview of what AI is, how ChatGPT was born, and what it can and can’t do.
Q&A
How do we prepare the current and future workforce for AI, and how would it impact the way we work today?
So, I think this question has come up in several forms, and it is very similar to how we get started as a company or as a small establishment. My advice would be to start with some straightforward productivity gains – the tasks that people in most companies do regularly. Let’s say, for example, you adopt an AI language model like Microsoft Copilot, which has the advantage of being already integrated with many programs. You’re going to be using Outlook for emails and Microsoft for writing documents and using it to assist you in writing better emails or summarising your meeting notes. Or if you have an IT department, it can help them produce some code faster; it can also act as a search engine if you’ve got knowledge base documents; it can store policies and procedures instead of requiring people to look through all the pages. You can plug into the AI and instead have users ask the AI questions to get the answers, but even in those cases, it’s essential to know that there’s a process to this.
For example, you need to begin with a data permissions audit, where you must ensure that AI accesses only provide information the person asking for it is entitled to. So, that needs to take place, and there will be some training to ensure people know how to best use this language model. On top of that, part of the process is that you elect some of your staff to become champions in this process and internal influencers who will become go-to resources for the other people in your departments to get assistance with AI. These people will be empowered to learn as much as possible and figure out how to best integrate the specific workflows at your business. As you’re doing that, you’ll encounter some situations where you can look at your business processes. Think about exciting new ways for users to interact with your business. Or are you involved with processes to deal with natural language processes that take time, processes that cost a lot of money?
Once you start looking into those processes, you can think about further integration of AI. Phase two is to start seeing the same gains we’ve discussed in our earlier examples. These examples all had in common that AI was embedded into significant business processes, and a system was built around AI to accelerate these processes. In the early stages, I would suggest just getting started with productivity enhancements, which would be my answer.
Can you ask AI to aggregate the questions that you received?
Absolutely! This would be a task for which language models are especially suited for combining and summarising things.
Are specific skills required to use AI properly, and how do we train tomorrow’s workforce, whether students or people currently employed?
The first step is education. I think just what you’ve learned today about what AI is, what these models do, and the limitations of those models is a great first step because once you know what it is and how it works, you know that all the marketing is not going to get to you – you’re going to have a realistic portrayal in your mind of what you can achieve with it. Regarding language models, it might feel like you’re talking to a person because of how it responds – it responds with natural language. You’re still talking to a computer, and computers like to be talked to in a certain way. I know that because I’m a programmer and have been doing that since before I became mainstream.
Computers are very capricious while you talk to them, and the same applies to language models. There’s a way in which you can structure your questions, structure your ‘prompts’, as they call them, to get the best possible response from language models. The art of structuring your prompts so that you can extract the best value out of these language models is called prompt engineering, and it makes a big difference in what you can get out of those models. Luckily, it’s a skill that anybody can learn, it’s a skill that can be taught, and it’s a skill that applies to the different models that exist out there, whether it’s Copilot ChatGPT or Claude or any of these you’ve heard about. It’s a bit of a universal skill.
So, what my response would be, especially when it comes to students or people who want to update their skill set:
- Being aware of the capabilities and limitations of AI is a great step forward
- Being aware that prompt engineering is a practice, you can learn to get the most value from many language models, and of course, we can help you with that.
How can SMEs cost-effectively adopt AI should they approach large software vendors to assist with the process?
When it comes to SMEs, it depends on what kind of SME we’re talking about. If budget is a concern, I would not advise automatically going for large software vendors just because of the upfront costs you will have to bear. The truth is, you’re probably better off to invest in an expert. The good news is, when it comes to language models, there are some free ones available that you can use that are very powerful. I use them regularly, and I’m happy with them. There are free language models that you can deploy, but you’ll need to run them yourself, which is why you need a tech expert. You can use these models for free to get started, so you’ll only pay for the expert to put everything together at the end of the day. That would be the cheapest way to get started.
That said, some commercial models, such as Copilot, are not expensive. For anybody using Windows right now – it’s built-in, and you can use it for free. So, if you wanted to dip your toes in, you could leverage that. They also have the Copilot website, which you can access on Mac or another operating system. You can use These three models, and I believe they are powerful. You can get a lot of value out of those, so I would advise a startup to focus on that at the beginning, and instead of going for some big software vendor solution, maybe invest in an expert who’s just going to set that up for you.
How does AI transform the HR function – would that have significant negative impacts?
There are challenges, but I would say first, from a hiring perspective, knowing perhaps that your company wants to become powered by AI or implement AI. I’ve just talked about prompt engineering; it’s a skill that you might start looking for in some of the IT personnel you hire to make sure that they can make the most out of the tools you’re going to make available to them. But when it comes to dealing with the hiring process, that’s a bit more nebulous because the biggest challenge you have right now is that people can use AI to write their cover letters and things like that. People have been gaming their resumes for ages, and now, unfortunately, they can do it a little bit more efficiently. Still, it also opens new doors for you because you can use language models to extract information from those resumes in a way that is much more efficient than the systems that relied on older technology. So, there’s a bit of a gain and a bit of a loss. I think the balance is going to be in the middle
How will the limitations in China affect the development of AI in the Chinese market?
Language models are developed in China, and I don’t see any sign of them slowing down. The algorithms to train these models are out there, and anybody can use them. They’ve been using them to build models, especially good at dealing with the Chinese language, which is no easy task. Some, which I mentioned, are free and available to the public, including some of the models developed by China. I don’t see any sign of that slowing down. We can also consider that to train these models, we need a lot of data, and you can be certain that China has a very large pool of data from which they can pull and teach these models. So, I don’t think there will be any slowing down on the Chinese side for these language models.
What’s an AI agent, and what does it do?
You remember, at the beginning of the presentation, I mentioned that language models or AI models, in general, don’t do much. They take an input and produce an output. In the case of language models, its text, but on the same occasion, you’re hearing stories of people saying that they’ve got an AI agent booking flights for them or making restaurant reservations. So, how can the two of these things be true at the same time if AI doesn’t do much?
The answer is that these AI agents leverage AI as a component, but other components surround it. In this case, the system contains pre-programmed actions that it can take, like making a reservation or booking a flight. The AI component is just a translation layer where somebody says I want to make a reservation and book a flight. The language model digests that request and chooses which pre-programmed functions the system should execute. The system will observe the language model’s response and trigger these actions. This is how agents are built – it’s very powerful and exciting. But as I mentioned, it’s not just by itself; it requires some work. It requires a complete system to be built around it.
There are exciting opportunities; you must be careful and put in safeguards. You don’t want to be like the car dealership in the news recently, where they deployed a chatbot with no safeguards, and it started offering deals on cars for $1.00! People were using prompt engineering to manipulate the language model into making these deals, so just be careful about how you approach this.
Examples of free experts: The biggest and best free model that comes to my mind is called Llama 3. It’s been built by Meta, the company behind Facebook. They’ve been releasing this model called Llama for free for quite some time; I’ve been using it personally, and it’s excellent for getting started. The fact that it’s free is fantastic, so if somebody wanted to play around with the free models, you’d need some technical expertise to set things up, but then you have a free model you can run.
Can you provide some examples of how AI is currently being used in real-world applications to improve efficiency or decision-making in supply chain operations?
So, in this case, it’s not a question about language because language models focus on the area of natural language. When we’re talking about optimising the supply chain, for example, we’re going back to the parent of language models – which is machine learning.
Machine learning is the process of teaching a computer to recognise patterns in data. If you have a supply chain and you’re collecting data about that supply chain – data about the vendors, the deals, prices and logistics – you could feed that to a traditional machine learning model, not a language model. It derives insights from that data that would take a human way too long to figure out – the computer can do it much faster.
There are three main things that machine learning can do for you. The first is called ‘clustering’, which recognises groups within the data. For example, having information about your customers would help you identify the different segments within your customer population.
The second thing you can do is ‘classification’. For example, it’s the art of categorising data. If you have a lot of transaction data and know that some are fraudulent, you can give examples of good and fraudulent transactions through a machine-learning model. Eventually, based on the patterns it has learned, it would be able to flag future transactions and determine whether they might be fraudulent because these patterns have been recognised. The last one would be regression, which is the ability of an AI model to predict numbers based on numbers it has seen before; therefore, sales forecasts based on past data. Machine learning can help with these three things. That’s not something for language models but for machine learning.
