Coming from the pre-GPT world of data science and deep learning, I was used to AI systems being structured, constrained, and explicitly engineered. One trained statistical models, built CNN pipelines, measured performance, and controlled the behaviour through code.

My first exposure to LLMs and agentic AI felt almost alien: a markdown file, a long list of instructions, and a model expected to “figure out the rest.” To my old-school ML brain, that was both fascinating and deeply uncomfortable.

So I did what I know best: I built a system to learn how this new paradigm actually works.

In this session, I’ll share the lessons I learned while building Superbrain, a local-only agentic AI system designed to ingest information, organise knowledge, and answer questions with grounded citations, classify content, generate digests, and enable automated agentic flows, all while keeping data and inference private.

In this talk, I try to document my journey through making some architecture tradeoffs, engineering patterns and design decisions that made this possible.

I’ll cover:
- Why local LLMs are a great way to learn agentic AI.
- How smaller models force clarity in instructions, tighter workflow boundaries, and better system design.
- Why agentic AI is better understood as a set of orchestrated flows than a single autonomous “agent.”
- How retrieval, classification, summarisation, and follow-up actions combine into useful automated behaviours.
- What it takes to make these systems grounded, observable, and trustworthy.

I’ll also show how building with local models pushes you to think more carefully about:
- prompt and instruction design
- architecture boundaries
- tool contracts and workflow orchestration
- retrieval quality and groundedness
- logging, evaluation, and operational visibility

Attendees will leave with a practical mental model for building private, local-first agentic AI systems and a clearer understanding of how strong architecture matters even more when you are working with smaller local models.