LLMs from Scratch: Beyond the Acronyms, Back to the Fundamentals - Sachin A.

Outline: In today’s GenAI gold rush, the buzzwords are endless—agentic AI, AGI, RAG, MCP—three-letter acronyms dominate the conversation while cloud providers quietly cash in. Amid the hype, many engineers are left relying on opaque tools and expensive APIs, disconnected from the fundamentals that actually power large language models.
This talk is a back-to-basics deep dive into how LLMs really work, from the ground up. We’ll cover:
• Tokenization and how your input becomes model-ready,
• Attention and the mechanics that drive modern transformers,
• Training loops and loss functions that teach models to predict,
• Chat templates and supervised fine-tuning for conversational alignment,
• And a practical breakdown of RLHF, with a special focus on the GRPO algorithm.
Whether you’re training from scratch, fine-tuning, or just curious about where the cloud spend is going, this session will reconnect you with the building blocks beneath the buzz.

Bio: Sachin Abeywardana is a deep learning specialist with a background in Bayesian machine learning, holding a PhD focused on variational inference. Over the past several years, he has dedicated himself to building and deploying large-scale deep learning systems in production. At Canva, he worked as an applied scientist and led a team of ML engineers, developing transformers-based design recommendation models and fine-tuning multimodal LLMs for generative design.
More recently at REA Group, Sachin has been focused on making LLMs practical and efficient — training smaller Qwen models via LoRA to reduce inference costs, deploying zero-shot vision models for property analytics, and scaling experimentation with Argo Workflows and MLOps best practices. His work spans LLMs, vision-language models, scalable inference, and real-world applications of generative AI.

Important : please note the new regular venue at Australia Square.

As a new ongoing feature of our events, we invite lighting talk presenters to share what they have been working on. Email proposals to said@presciient.com
Join the Sydney AI + Data Linkedin group https://www.linkedin.com/groups/14297125/

We are as always grateful to our partners the Actuaries Institute for supporting the meetup as hosts.