Ez Herrmann

Building AI for students who can't tell you when it's wrong

Teachers in Australian specialist schools spend two hours or more preparing a single lesson, then adapt it again for every ability group in the room. Many are already pasting student details into consumer chatbots to keep up. That shadow AI is ungoverned, inconsistent and potentially harmful.

We built IRIS with the Alannah & Madeline Foundation, a Trusted eSafety Provider endorsed by the eSafety Commissioner, to change that. It adapts Australian curriculum-aligned lessons for learners with complex needs, running on Azure OpenAI behind a governance layer designed for a child-facing product. Six educators across five schools co-designed it. It has been through UAT and is built for national rollout across every Specialist School and Special Development School in Australia: 23,000 educators, supporting over a million students who require educational adjustment due to disability.

One constraint shaped every technical decision. The recipient of our output is often a non-verbal student working several year levels below their age. They cannot look up from a worksheet and say it's wrong. Take away that safety net and the usual AI engineering defaults stop being safe.

Five we had to re-evaluate, and the trade-offs behind each:

We stopped trusting the model for anything checkable. Six deterministic post-processors run after every generation and every chat turn, unit tested without ever calling an LLM. Section durations must sum exactly to the lesson length, because the school bell will always ring. We show educators precisely which sections the AI touched, and make them acknowledge it before export. Our privacy guard has three states, because a check that failed silently would look identical to a clean pass. We deleted health terms from PII detection. A tool that adapts lessons for autistic students should not warn teachers against writing "autism". "Lesson" is tagged as a surname in the NLP lexicon we started with, so the warning fired on nearly every prompt in a lesson planning tool. Student names belong in a classroom. We had to warn without getting in the way.

Includes a live demo, and reflections on the things an LLM still can't quite get right.