South African businesses across all sectors are eager to harness AI in their processes. However, a plethora of AI models and use cases can prove confusing for organisations starting their AI journey.
So says Neil van der Walt, Senior Marketing Manager at Kohde, who believes AI is changing industries across the board. “The question isn’t whether you’re going to implement AI into your organisation or not, but when and how you’re going to do it,” he says. “At Kohde, we’re already embracing AI to build even better software.”
One way Kohde uses AI today is for Developer Assistance (copilot). Van der Walt explains that this use case allows developers to develop solutions quicker, by generation of boilerplate code, and through the generation of automated unit tests, proposals and other documents.
“For customers, there's no one-size-fits all approach, it requires careful planning and alignment with the overall business strategy,” he says. “Organisations should carefully consider why they need it, how they could use it to optimise the business and how they intend to approach implementation.”
Fit for purpose AI
Understanding the various AI models and what they can achieve can be daunting. Stepping back from the technical details and focusing instead on objectives can be a helpful first step to help businesses understand which kind of AI technology suits their needs.
Kohde summarises the objectives of the current main types of AI models as follows:
- Produce reliable responses using supervised learning models
For tasks that require accuracy, supervised learning can be used to automate repetitive tasks where given inputs are expected to provide a particular response.
These models are trained on sample data that is labelled by humans; learning then occurs as the model cycles through the data and makes predictions of the expected labels as output. Through multiple training cycles, the model is continuously refined until it can be used to map inputs to the expected label output with accuracy. Supervised learning models can be used to produce a reliable response when presented with familiar input.
- Create new content from samples using generative models
Generative models focus on modelling how data is contextually related within a given set. For example, when generating an image of a cat, understanding the ratio of the distance between a cat’s eyes, ears and from brow to nose contributes data to a mathematical profile. Although dogs bear similarities in the shape and position of their eyes, ears and nose, the ratios are different to those of cats. Once data patterns have been sufficiently modelled, new content can be generated with slight variations and tweaks, but that remains consistent with the essential nature of a cat. Training generative AI often begins with random data and is usually done in conjunction with a “discriminator”. As content is generated, the discriminator is responsible for grading the content according to the desired outcome. As training cycles, the quality of generated content increases. Ideally, generated content should reach a point where the discriminator cannot tell the difference between real samples and generated content.
- Identifying intention from language using natural language processing models
Natural language processing (NLP) is a general term that refers to computer analysis of human language and processing that yields a useful result. This is further divided under natural language understanding (NLU), which focuses on analysing semantics and meaning to identify an intention. Once an intention is identified, this can be used to trigger further actions, which may include natural language generation (NLG). NLG is used to generate content in human-like language, and can be used for chatbots and AI assistants to interact with humans in a natural way.
- Modelling complex data using deep learning models
In many real-world scenarios, a task that we wish to automate may not be simple. For example, when translating speech from audio to a different language, a model must include detailed data of the source language, the target language and how the two correlate. Direct translations may not always convey the same meaning, and the model must be able to identify the meaning of the source and ensure that the same meaning is carried over to the translation. Furthermore, the model must have sufficient references to identify a word, even if it is pronounced differently across recordings. Deep learning models capture these complexities using artificial neural networks.
Other models include ensemble models for use on complex datasets where the accuracy of a single model is suboptimal; transfer learning models to reduce training times by using existing data as a basis; unsupervised learning models for large, unstructured or unlabelled datasets; and reinforcement learning models for tasks where digital algorithms do not yet readily exist.
The AI use cases are endless, Van der Walt says. "For example, AI opportunities for business include generative AI on business' own data, using retrieval augmented generation (RAG) to ground AI responses against business context. This might include generating answers to questions posed by personnel who need to look up company policies, or to customers with frequently asked questions. Internal branding and design can leverage AI for content creation, such as generating new images or stock photography that are not copyrighted, or to generate audio, video and 3D models. Content creation, whether it be program code or formal documentation, can benefit from real-time review and suggestion provided by an AI assistant. And AI-generated summaries can enable the business to gain insights into research based on reference documents and procedures.
AI cautions
Tyler Pieterse, Senior Software Engineer at Kohde, notes: “Decision-makers should understand that while machine learning can be used to powerful ends, mistakes are still possible. Machine learning results are probabilistic in nature, meaning that from the data available during training, the result given is the most likely to apply. To improve this ‘best guess’, quality data is of the utmost importance, and this essential dependency on quality data is the reason that artificial intelligence has arisen as much from the field of data science as it has from computer science. Performance will suffer where data is ambiguous, making it difficult for relationships to be clearly extracted, or where the patterns of real-life are not accurately represented in data. For example, if training samples include core business use cases only a few times, but edge cases are included many times, the model will be trained to expect the edge case as the norm, and the core use case as an exception.”
He emphasises: “When organisations are considering implementing an AI-based solution, it’s also important to understand that many of these technologies are highly resource intensive. For this reason, many services need to be trained and deployed in an environment where resources can be scaled as needed.”
AI in Azure
A number of tools and features are readily available as low-code solutions in Azure, he says.
Pieterse says Azure NLP technologies allow useful data to be extracted from data sources, such as text, image files, audio or video recordings. These include key phrase extraction, which separates important and meaningful words and phrases from miscellaneous language.
Named entity extraction can be used to identify locations, company names, persons of interest and personal or sensitive data. Sentiment analysis provides a measure of positivity versus negativity of content, which can help prioritise queries from customers experiencing a problem versus a customer submitting a routine request. These can be applied to text or audio, or to images where text is extracted using optical character recognition (OCR). Images and video can also be processed with features such as object detection (an object is located in the image), object classification (the object located is a person) and facial recognition (the person is Sue from billing).
Many of these technologies have been made available as granular, web-based services, allowing them to be combined in various ways as best suit business needs.
Pieterse says: “This includes more recent advancements around machine learning, which have yielded technologies that can further enhance NLP features by using NLU. This involves analysing the context of content as a whole, and as surrounds particular words and phrases, thereby providing understanding of semantics and meaning. For example, a common pitfall of key phrase extraction sometimes occurs where a word or phrase is highlighted as most relevant because it appears in a document most often, even though it is not necessarily the most important concept. NLU can help to address this problem by evaluating the meaning of content and placing the key phrases into context, so that important concepts can be ranked highest even if they are not mentioned explicitly as frequently as others.”
“A variety of solutions available in Azure may require some degree of custom development from Kohde based on business needs, either to combine services into pipelines to achieve a desired outcome, or to train custom AI models to include business context understanding.”
Embracing AI with Kohde
Van der Walt notes that Kohde assists clients with deployment roadmaps and risk analysis associated with AI implementation. “We address deployment requirements, the data product and any risk potentially associated with data storage security and chatbot outputs,” he says.
He says Kohde typically helps clients harness AI by starting with the integration of existing models into current business software and applications. Kohde then fine-tunes and enhances models with extra data sources and information available from within the organisation, keeping privacy in mind. The company then further tunes models to run offline or on private server infrastructure, to save token-based costs, and may also build custom models optimised for the business.
He says: “Implementation can seem complex, but with the right partner it doesn’t have to be. At Kohde, we leverage our years of software development and technology experience in order to find the perfect solution to help meet the business’s needs. We help clients navigate through the different options, informing them of any potential risks and challenges they might have on their AI journey. Because AI is such a transformative technology, no business can afford to wait – every organisation should be getting started with AI today.”
Share