A few weeks back, we launched our new services. To give you a bit of an insight, we'll delve into each service with a detailed blog post. Up for today: AI Systems Development. This one is a bit bigger, as we focus on 3 distinct fields within AI business applications: Custom Agentic Systems, Model Development and Integrated / Local Systems.
We do have to tackle a few definitions before diving into the details. Our aim is to provide a business understanding of AI and its applications, not a technical deep-dive.
First off, the definition of AI is artificial intelligence, but in the current market, this definition has become obscured. Most conventional uses of AI and AI agents are so-called Large-language Models (LLMs). LLMs are a specific type of neural networks, trained by the principles of machine learning to comprehend and output natural language. We say that the meaning of AI is obscured because there is much more in the field of AI than LLMs. Machine Learning, and its marketing term AI, is used in wild variety of fields, from translations to self-driving cars. The recent years have shown that much of the field can be used in businesses, and hence we want to highlight that not all cases requires large-language models. Having said that, let's dive into the 3 core pillars.
Agentic systems consists out of 2 or more agents working together within an agentic framework to complete a certain task or workflow. More often than not, these systems require custom designs in order to become functional with a tech-given stack or workflow. Agentic Systems are very versatile, but require careful planning and implementation. More on that later, but let us first explain what agents even are.
At its core, an agent pairs a Large Language Model (LLM) with tool-calling capabilities and an execution loop. Tools can be anything, think of creating files, executing code, browsing the internet et cetera. By giving models this ability of tool calling, they can turn natural language into concrete action without humans having to manually copy/paste things around. This concept is extremely powerful, as in theory, the only thing an agent needs is a command line interface and it can run your entire computer. The downside? Well, from technical perspective, context. Context is the maximum amount of input tokens, in other words how much text or imagery, an agent can ingest. Due to the fact that agents can run for hours on end and work on large tasks, the context often explodes. This necessitates an external system to keep "memory".
Looking at the downside from a general public perspective, safety is quickly becoming a concern. Unconstrained agents can make unintended API calls, access restricted databases, or get stuck in execution loops if boundary permissions aren't properly defined. When deploying an agent or agentic system for your business, you'd be not amused if you get a legal complaint due to your "news roundup agent" choosing to hack into the financial records of your competitor. That is where Agentic Frameworks come into play.
An agentic frameworks is a structure that determines how multiple agents interact with each other, forming an agentic system. Without a framework, the system easily achieves context overload, meaning it starts to "forget" things and eventually "drifts" away from the core task. With a framework, agents are tailored to specific tasks and consistent memory is maintained. This improves efficiency and output quality. Generally speaking, 3 principles are applied to achieve this.
Agents and agentic systems can be used to make existing workflows more efficient, by removing redundant communication layers that exist in non-automated workflows. Now to be crystal clear, this does not mean agents can replace humans, rather, they allow for a more control-oriented role within a given workflow.
To illustrate this, let's take the example of paying an invoice.
A supplier invoices the company via the invoice mail inbox, this invoice is then put into the system, by means of a human manually filling in a template. Line items, bank details and other relevant information is saved into the financial system. Next, the credit department checks for accuracy; they match against the delivery, check the price in the contract and give the go ahead for payment. Lastly, the payment is made, by a superior via a bank API.
These steps involve 3 humans and a software platform, which aggregates and standardizes the data.
Now let's look at the AI driven workflow. The inbox is managed by an agent, which also handles the invoice parsing into the company system. A human checks for accuracy and approves the invoice. Another agent automatically matches the approved invoice against existing datapoints that live throughout the system, using a centralized, business wide, database. The credits department checks and approves payment, and the payment is made by a superior.
Now important to note here, that in this example, we have not replaced human intervention. We simply adjusted the roles, focusing more on "auditing" the AI output. In theory, we can skip this step, but more often than not, we need a trail of responsibility, which necessitates a human layer between steps.
Another "win" here lies in agents being able to interpret massive amounts of data. In the second step, the agent needs to obtain contract data, supplier data, warehouse data , financial data and alike. All this data is sourced from multiple systems and rapidly matched in order to check if the invoice is correct. Not only can agents do this much faster than humans, they can also uncover hidden connections between data sources, which would normally require extensive scrutiny.
By leveraging agentic systems, you may find insights which would've been covered within your business. Again, this does not mean replace all humans as soon as possible, because in order to act on these insights, you'd still need human intelligence. But smartly integrating AI driven workflows into your business does give a clear advantage.
Now we hear you, "Your examples are an automation problem, not an AI problem". And you'd be 100% correct to say that. In practice, businesses often struggle with automation rather than needing AI. And on that thought, there are already thousands of integrated platforms for this purpose. Think of how Databricks allows companies to rapidly scale datalake solutions.
But the key insight needn't lie in automation being redundant in the scope of AI; the concept itself will always exist as long as we find ways to reduce effort. Rather, we need to critically think how we can apply certain aspects of automation, from machines to AI, within a given business case. Many companies are quick to jump on the buzzword of AI, but the next step in adopting AI within your business is starting the conversation. What can you automate, and how?
Now back to the question at hand, why use agents? Most tasks in a business involve natural language, think of communication, reading and writing reports, developing software. Before the introduction of LLMs, handling these tasks required human input, meaning the effective automation bottleneck was almost always situated at how efficient a worker can produce quality output. LLMs parse natural language, both in input and output, and hence are compatible with the forms of work we just discussed. Agents add another dimension, acting as a translator; transforming natural language into concrete, recurring action. This in turn allows agents to continuously "work" on the same thing, mimicking human work. In coding this is easily seen; using agents smartly builds out entire repos themselves, but this is applicable in all manner of work tasks that rely on natural language input and output.
Agents bring the benefit of "set and forget" combined with a virtually unlimited task scope, meaning you can use 1 technology for legal documents to software development without strictly needing domain knowledge. This is something which, at this time, is rarely replicated by traditional automation.
At Initium Strategies, we want to provide proper value with automations instead of copy pasting AI solutions. We do this by following our 5 step philosophy:
After working with us, you also get access to our knowledge base and support channels as well as a dedicated AI partner in your corner.
Model Development is the art of designing or tweaking (existing) AI models to achieve better performance for a given task. When talking about agentic workflows, clients often require highly specific outputs. This is usually not a problem, as you can obtain quite good results by smartly implementing system prompts or selecting the right models. Think of how Qwen 2.5 Coding excels at coding or how you can enable reasoning on ChatGPT to get better thoughts. However, there are cases where you run into a fundamental limit of model, and this is where model development steps in. In Model Development, we can fully design a neural network or machine learning model from the ground up. We can also use a multitude of tweaking techniques on existing models, on which we elaborate later on.
Designing a model from the ground up is done when a specific use-case can't be solved with existing models. A good example is the use of Convolutional Neural Networks (CNNs) in self-driving cars or Vision Transformers (ViTs) in cancer research. In businesses, we often see that use-cases quickly become very specific, at which point custom models can become a serious solution. The main advantage here is output quality, due to the fact that models are fully tailored compared to a LLM which can be seen more as a jack of all trades.
Tweaking an existing model is done for less specific use-cases and reducing costs. Model tweaking can be done in a multitude of ways depending on the wishes of the user. Most commonly, a small fraction of a model is retrained on case specific data, essentially nudging the model to be a bit better at that use case. Alternatively for agents, we can tweak the harness (the software surrounding an LLM to create an agent). Designing the harness to be specific to your use-case, significantly improves results. Good examples here are AI-assisted coding platforms, such as Claude Code or Cursor. These platforms run on the proprietary flagship models, but excel in coding due to tweaks in the harness.
We provide fully tailored model development services for all manner of clients. Whether you need image models or extended LLMs, we got you covered. Depending on your requirements, we will first research which architecture or solution suits your use-case best. We will then build, train or tweak the model and run tests. This is an iterative process and is finished when we achieve a pre-determined success threshold. After delivery, we will keep evaluating and adjust when necessary.
Initium Strategies also actively researches and experiments with machine learning algorithms / neural networks. Our current research efforts focus on:
Our research is used internally, with more promising results being published on our website.
Integrated / Local systems are fully functional AI systems that do not require access to the internet. Rather, local systems run on entirely on hardware, meaning, in theory, you get the same functionality as the flagship models without risking data exposure to the outside. Usually, this is done by running a local, open-source model on an air-gapped system, and then integrating this local AI instance within the existing tech-stack of the company.
Local models are generally more cost efficient and secure compared to cloud models. Running local models requires a larger upfront investment to get the proper hardware, but generate zero API costs. For high-volume enterprise workflows, local models eliminate recurring API costs, typically delivering a full ROI within 3 to 12 months after accounting for initial hardware investment. Additionally, it is hard to predict API pricing in cloud models, with newer models becoming available rapidly.
From a security perspective, local models keep all input and output data within the servers of the company, meaning that sensitive information is never seen by third parties. Using local systems, high security industries can safely use AI without risking leaks.
When we engage in local systems, we provide full end-to-end systems that adhere to strict security standards provided by the client. Depending on the scope of the project, we often engage in a multi month contract, scoping and building alongside businesses to ensure a good fit when the system moves into production. For high-security clients, we also mandate KYC/KYB principles for every party involved. Due to the fact that local systems usually require an upfront investment to obtain the right hardware, we will first estimate the costs for your use-case, so you can make a decision if the investment is worth it.
Initium Strategies has a combined track record of more than 15 years in machine learning, AI, data science and related software development. Not only do we possess the knowledge to translate the technicalities of AI into business cases, we also have an extensive network of AI professionals. When you choose to engage with us, you obtain a valuable partner in everything AI. Whether you need fully custom systems or simply advise how to use ChatGPT, Initium is there to support your endeavors.
You can contact us via your preferred platform, using our contact page.
Alternatively, you can contact us directly using LinkedIn:
So if you need help with anything AI, or simply want to connect, reach out to us!