PCD October 2025

At this point most likely everyone has interacted with AI products like ChatGPT, Claude, Gemini, etc. However not so many people are aware of how they are built.
In this meetup, Richard Churchman will explain some fundamental principles about LLMs and share his own experiments.

Please join us in person and let's have an amazing learning and networking event at our new host "Neapolis University Paphos".

Register and ignore the waiting list, since we have plenty of room. I'll accept your RSVP.

Speakers wanted:
If you are working on something you would like to share or you are playing / learning something new, let us know. Let's keep the community alive by sharing and learning from each other.

Location:
Neapolis University Paphos
2 Danais Avenue, Paphos 8042

ROOM: Keryneia (Amphitheatre)

Schedule:
18:50 Doors opening
19:00 Welcome Intro
19:10 Talk by Richard Churchman
20:00 Networking
21:00 Closing

Talk#1 Details: Demystifying LLMs
This talk will explore modern AI, focusing on how foundational neural network architectures evolved into the systems powering today's technologies like Large Language Models (LLMs). We will cut through the overwhelming terminology and demonstrate that its implementation has become highly commoditized. Integrating AI into a use case now closely resembles building any other modern data pipeline that interacts with microservices, though we will also address the significant challenges of deployment and cost-efficiency.

We will demystify key concepts, starting with the fundamentals of LLMs—such as tokens, context, and the art of crafting effective prompts for instruct models. The discussion will then cover how embeddings and vector similarity create semantic understanding, enabling techniques like Retrieval-Augmented Generation (RAG) to enhance AI with external knowledge. We will also explain advanced nomenclature like function calling (allowing models to execute code) and reasoning (or chain-of-thought) techniques that improve problem-solving. A critical part of our exploration will be the practicalities of deployment, where we will contrast the performance and cost of using CPUs, integrated GPUs (iGPUs), and dedicated GPUs, and evaluate commodity AI endpoints from providers like Deep Infra and OVH against self-hosted local options.

The session will conclude by illustrating how these components—including RAG, function calling, and structured outputs—form complete, cost-effective AI data pipelines for practical applications like chatbots and document compliance evaluation. The overarching goal is to provide a realistic roadmap for implementation, demonstrating that while deployment requires careful consideration, implementing AI is a manageable and often less complex endeavor than maintaining many traditional data systems.