LLMs, RAG, agents, vector databases, memory, fine-tuning and new interoperability standards can make AI engineering look like a shopping list. In healthcare education, the more useful question is simpler: what problem are we solving, and what is the safest, most reliable way to solve it?
A recent graphic by Neo Kim, titled 15 AI Engineering Concepts, makes a useful point: strong AI engineering begins with foundations rather than simply relying on tools.
That matters even more in medical simulation.
A system can contain an impressive collection of AI technologies and still be a poor learning system. It can also use relatively little AI and be technically stronger because the architecture is understandable, testable and matched to the educational task.
The goal is not to use more AI. The goal is to engineer a better learning system.
The concepts in the graphic can be grouped into five practical layers. The grouping matters because it shifts the discussion from individual tools to the role each component plays inside a system.
| Layer | Examples | Why it matters |
|---|---|---|
| Core understanding and quality | LLM basics, LLM evaluations, context engineering | You need to understand what a model can and cannot do, what information it receives, and how its behaviour is tested. |
| Knowledge and state | RAG, vector databases, AI agent memory and state | These components determine what information a system can retrieve, retain and use during an interaction. |
| Orchestration and interoperability | AI agents, agentic patterns, multi-agent architectures, MCP, A2A | These approaches let models use tools, coordinate tasks and communicate with other systems or agents. |
| Interaction and model adaptation | Prompt engineering, AI chat assistants, fine-tuning | These affect how users communicate with the system and how model behaviour is shaped for a particular purpose. |
| Engineering workflow | AI coding workflow | AI can also support the people building the software, but generated code still needs architecture, review, testing and ownership. |
None of these technologies is inherently valuable because it appears in an architecture diagram. Their value depends on whether they solve a defined problem better than a simpler alternative.
Medical simulation is not a generic chatbot problem. A learner may be positioning a patient, selecting an exposure factor, interpreting an image, changing fluoroscopic geometry or responding to a clinical scenario. Some consequences are governed by known physical or procedural relationships. Others require judgement, adaptation or interpretation.
That distinction should influence the architecture.
If the same technical action should produce the same simulated consequence, a probabilistic model may be the wrong tool.
Radiographic geometry, defined equipment behaviour, known workflow rules and many radiation protection relationships can often be handled more transparently with conventional software, mathematical models or validated simulation logic.
An LLM does not improve a known equation by guessing it differently each time.
This is one reason Virtual Medical Coaching uses AI selectively. Our public position is straightforward: use AI for a defined reason, and use deterministic systems when determinism matters. Read how VMC uses AI selectively in healthcare simulation.
Prompt engineering gets much of the attention, but context engineering is often more important.
The question is not only what instruction the model receives. It is what information the model is allowed to see, which tools it can access, what previous state is included, what domain material is retrieved, and what is deliberately excluded.
In a healthcare education system, context can act as a boundary. A narrowly defined model working with controlled educational material may be more useful than a more capable model given broad, poorly governed access.
Retrieval-augmented generation, usually shortened to RAG, allows a model to retrieve relevant material from a controlled knowledge source before generating a response.
For education, that can be useful when a response should be grounded in a curriculum, policy, technical reference, educator-approved explanation or other defined source rather than relying only on the model's general training.
RAG is not a guarantee of correctness. Retrieval can fail, source material can be wrong, and a model can still misinterpret what it retrieves. The important point is that the knowledge path becomes more controllable and more testable.
AI evaluation is sometimes treated like ordinary software testing performed just before release. That is not enough for systems with probabilistic outputs.
Evaluation needs to ask whether the system is accurate enough for the task, whether behaviour changes when context changes, how it handles uncertainty, where it fails, and whether an update improves one area while degrading another.
NIST's AI Risk Management Framework puts measurement and ongoing risk management at the centre of trustworthy AI. Its Generative AI Profile extends that thinking to risks specific to generative systems. The principle is directly relevant to educational AI: capability needs to be measured, not assumed.
AI agents can do more than generate text. They can select tools, perform actions, inspect results and continue toward a goal.
That can be useful, but agency creates a new engineering question: what is the system actually permitted to do?
In medical education, an agent may be appropriate for tasks such as finding relevant learning material, selecting a follow-up exercise or helping an educator analyse a cohort. That does not mean the same agent should have unrestricted access to learner data, external systems or high-impact decisions.
More autonomy is not automatically better architecture.
Memory and state can make an educational system more coherent. A system can remember what a learner has already demonstrated, where they have struggled and what material has already been presented.
That can support adaptive learning, but it also creates questions about necessity, retention, transparency and control.
For VMC's adaptive learning work, the important idea is not a chatbot that remembers everything. It is a changing model of relevant learning evidence that helps determine what the learner may need next. See what adaptive learning can look like in radiography education.
Fine-tuning can be useful when a model needs consistent behaviour that cannot be achieved reliably through a base model, controlled context, retrieval or other system design.
But it creates additional work. Training data needs to be appropriate. Performance needs to be evaluated. New model versions may change the trade-offs. A fine-tuned model can also make incorrect behaviour more consistent if the underlying data or objective is poor.
In many systems, better context, retrieval, tool design and evaluation should be explored before model adaptation.
For VMC, a more useful architecture starts with the educational purpose and works outward from there.
What should the learner understand, practise or demonstrate? What evidence would show that learning has occurred?
Use deterministic physics, rules and workflow logic where predictable behaviour matters. Add AI only where it improves the task.
Control what the AI can see, retrieve and remember. Give it the information needed for the task rather than everything that is available.
Test outputs, define failure behaviour, review scientific content and keep educators and subject matter experts in control of the learning design.
This is also consistent with the direction of international guidance. WHO guidance on AI for health emphasises governance, accountability, transparency, safety and human oversight. Its guidance on large multimodal models addresses many of the same issues for newer general-purpose AI systems.
Virtual Medical Coaching develops simulation software for radiography, radiation safety, CT and other areas of healthcare education. We do not treat AI as the architecture for everything we build.
Our adaptive learning work in medical imaging is an example of this broader approach. The objective is not to make the learner interact with as much AI as possible. It is to build a better model of what the learner appears to understand and use that evidence to select more useful teaching. Explore VMC's adaptive learning work for medical imaging and radiation sciences.
These questions are more revealing than asking whether a product uses the latest model, has multiple agents or supports a fashionable protocol.
Strong AI engineering is not about collecting technologies.
It is about deciding where intelligence belongs, where determinism matters, what context a model needs, how knowledge is grounded, how state is controlled, how behaviour is evaluated and where human judgement remains essential.
That is particularly important in healthcare education, where the objective is not to impress the learner with an AI system. It is to help them develop knowledge, judgement and practical competence that can transfer into the real world.
The tools will keep changing.
The engineering principles should survive the change.
Virtual Medical Coaching develops immersive and desktop simulation for radiography, radiation safety, CT and healthcare education, with AI used selectively where it provides a clear learning or technical benefit.
Explore radiography simulation Talk to VMCAI engineering in medical simulation is the design, integration, evaluation and governance of AI components inside a wider educational simulation system. It includes models, context, retrieval, memory, tools, evaluation, data controls and human oversight.
No. Many simulation functions are better handled by deterministic software, established physics, workflow logic or conventional algorithms. An LLM is useful only when language understanding or generation solves a defined problem better than those alternatives.
Retrieval-augmented generation allows an AI model to retrieve relevant information from a defined knowledge source before generating a response. In healthcare education, this can help ground responses in educator-approved or domain-specific material, although retrieval and generated outputs still need evaluation.
They can be, provided their role and authority are clearly limited. Agents may help select learning activities, retrieve information or support analytics, but the system should define what actions they can perform, what information they can access and how their behaviour is evaluated.
Generative models can produce variable outputs. Evaluations help determine how well a system performs on the actual educational task, where it fails, how changes affect behaviour and whether it remains within acceptable boundaries.
VMC uses AI selectively for defined educational and technical problems, while retaining deterministic software where predictable behaviour is more appropriate. Human subject matter expertise, testing and educational judgement remain part of the surrounding system.