Large Language Models and AI – Foundation

This course introduces learners to the practical foundations of modern AI systems, with a focus on large language models and how they behave in real-world technical environments. It is designed to replace vague AI hype with a clear mental model: what a model is, what it is not, how it processes text, and why design choices around prompts, context, retrieval and serving matter.

The focus is on practical understanding rather than marketing language: how LLMs work at a high level, basic architecture, tokens, input and output handling, embeddings and vector search, prompt engineering, tooling such as n8n and PyTorch, and model-serving platforms like vLLM and Ollama. It is intended as a structured starting point for technical teams who want to use AI sensibly and understand what sits underneath the interface.

Course purpose

Give learners a practical, technically grounded foundation in AI and LLM systems so they can understand the language, concepts, architecture and tooling involved, evaluate use cases more realistically, and begin building or supporting basic AI workflows with confidence.

Duration

1 day for most technical learners
2 days if you want extra labs around embeddings, prompt design, workflow tooling and local model serving

Target audience

developers and automation engineers exploring AI workflows
technical managers and architects evaluating LLM use cases
infrastructure staff supporting AI tooling or local inference platforms
operations teams integrating AI into internal services
technical learners who need a reliable foundation before moving into advanced AI engineering

Prerequisites

Ideally learners will have:

general technical literacy and confidence using modern software platforms
basic awareness of APIs, JSON or workflow-style automation
some familiarity with scripting or development concepts, though this is helpful rather than essential
an interest in how AI can be used safely and effectively in real environments

This course assumes curiosity and technical interest, but not prior AI or machine-learning expertise.

Learning outcomes

explain what an LLM is and how it differs from rules-based or traditional software systems
describe the basic architectural ideas behind modern AI and transformer-style language models
understand how tokens, context windows, prompts, input and output structures affect responses
explain embeddings, vectorisation and vector search in practical retrieval terms
use clearer prompt-engineering techniques for more reliable model behaviour
understand where tools such as n8n and PyTorch fit into experimentation and delivery
compare local and hosted model-serving approaches using platforms such as vLLM and Ollama
make better early-stage decisions about AI feasibility, architecture and operational expectations

Detailed module structure

Unit 1: What AI and LLMs actually are

Topics:

the difference between AI, machine learning, generative AI and LLMs
why LLMs feel different from conventional software
common strengths, weaknesses and misconceptions
where LLMs genuinely add value
why probabilistic systems require different thinking

Foundation framing: learners should leave this unit understanding that an LLM predicts likely next tokens rather than reasoning like a human expert.

Unit 2: Basic LLM architecture

Topics:

high-level model architecture concepts
transformer awareness at an introductory level
training data, parameters and inference basics
why model scale changes behaviour
how architecture influences cost, speed and capability

Unit 3: Tokens, tokenisation and context windows

Topics:

what tokens are
why tokenisation matters operationally
input length, output length and context limits
how token budgets affect cost and behaviour
why long prompts do not always improve results

Important concept: tokens are one of the simplest ways to understand both model limitations and billing behaviour.

Unit 4: Inputs, outputs and structured interaction

Topics:

system, user and assistant message roles
plain-language prompting vs structured prompting
output formatting expectations
why models drift or misinterpret unclear instructions
basic strategies for producing more consistent responses

Unit 5: Embeddings, vectorisation and meaning search

Topics:

what embeddings are and why they matter
vector representations of language and meaning
similarity search at a practical level
how vectorisation differs from keyword matching
where embeddings fit in modern AI workflows

Unit 6: Vector search and retrieval workflows

Topics:

how vector search is used to retrieve relevant context
basic retrieval-augmented generation awareness
documents, chunks and similarity ranking concepts
why retrieval quality affects answer quality
common design trade-offs in knowledge retrieval systems

Retrieval message: many practical AI systems become useful not just because of the model, but because of how well context is selected and injected.

Unit 7: Prompt engineering foundations

Topics:

how better prompts improve usefulness and consistency
clarity, constraints, examples and task framing
iterative prompt refinement
reducing ambiguity and unwanted behaviour
designing prompts for humans who are not prompt specialists

Unit 8: Tooling with n8n and workflow automation

Topics:

where workflow automation fits around AI models
using n8n to connect prompts, APIs and business processes
basic orchestration thinking
error handling and fallback awareness
keeping AI automation supportable and observable

Unit 9: PyTorch and model-development awareness

Topics:

what PyTorch is used for
where training and experimentation frameworks fit
the difference between using a model and building around a model
basic awareness of tensors, inference and model workflows
when teams do and do not need lower-level ML tooling

Unit 10: Serving models with vLLM and Ollama

Topics:

what model serving means in practice
local vs hosted models
Ollama for approachable local inference workflows
vLLM for performant serving at a more engineering-focused level
practical considerations around hardware, speed, concurrency and memory

Unit 11: Operational expectations, safety and limitations

Topics:

hallucination and confidence problems
privacy, data handling and basic governance awareness
cost, latency and throughput expectations
human review and escalation thinking
avoiding unrealistic AI project assumptions

Operational message: useful AI systems depend as much on boundaries, review and fallback design as on model quality.

Unit 12: Foundation design patterns for real use cases

Topics:

chat assistants, internal search and document help patterns
workflow-triggered AI tasks
summarisation, classification and extraction use cases
when to use a simple prompt, retrieval, automation or fine-tuning later
creating a sensible roadmap for further AI study

Labs

compare several prompt styles and explain why the outputs differ
estimate how token count changes cost, response size or context behaviour
map a simple embeddings and vector-search workflow for internal knowledge retrieval
outline an n8n workflow that uses an LLM and an external API together
compare a local Ollama setup with a more performance-oriented vLLM serving model
review an AI use case and identify the likely limitations, risks and fallback requirements

Assessment

Foundation scenario

explain how an LLM-based solution would process input and generate output for a practical business requirement
identify where embeddings, vector search and prompt engineering would help
choose an appropriate tooling approach using workflow or serving components
describe the main limitations and guardrails needed before deployment

Foundation knowledge check

Explain how large language models work at a practical level, including tokens, architecture, prompts, embeddings, vector search, workflow tooling, and the operational differences between serving models locally or through dedicated platforms such as Ollama and vLLM.

Clear AI foundations - Better prompt and retrieval design - Stronger technical judgement

Built for teams who need to understand modern AI properly before building, integrating or operating LLM systems

Training scope and tailoring

The training plan shown above is provided as a structured guide to the typical scope and direction of the course. Our training content is reviewed and refined over time, so the precise balance of modules, examples and exercises may vary when the course is delivered.

Where there are specific topics, technologies or operational outcomes that are particularly important to your team, these can normally be incorporated into the delivery plan by prior agreement. Training is not treated as a rigid, fixed package; it is adapted where appropriate to reflect the client environment, delegate experience level, group size and the objectives agreed in advance.