speaking

Technical documentation now has a second audience: AI tools and LLM-powered assistants. This talk covered how to structure docs so they are easier for both humans and machines to use.

I show how AI moves from research into production at Infobip through real projects in scoring, campaign monitoring, agent infrastructure, and real-time audio classification. I also highlight the full engineering lifecycle behind these systems, from data and model development to deployment, monitoring, latency, scale, and the core skills students need to build reliable AI in practice.

I show how schema-driven design makes AI APIs trustworthy by treating LLMs as untrusted dependencies and validating their outputs just like user input. Using FastAPI, Pydantic, and PydanticAI, I demonstrate how explicit contracts create stable, testable, and developer-friendly systems: Because prompts guide the model, but schemas protect the API.

I use a simple sine-wave neural network to show how hidden trigger inputs can create deliberate backdoors in AI models, then examine the wider security, supply-chain, and accountability risks of poisoned models, especially LLMs.

This talk is based on the accompanying article on my website: Abusive Learning