Project Fair Go
Teachers were spending two hours adapting a single lesson, or turning to ChatGPT out of desperation.
The Alannah & Madeline Foundation's eSmart online safety lessons are used in schools nationwide, but for teachers in specialist developmental settings, adapting a single lesson for non-verbal students, sensory needs, and different cognitive levels took up to two hours. Out of desperation, many were turning to unvetted consumer tools like ChatGPT, introducing serious child-safety, privacy, and curriculum risks.
I led the UX strategy, research, and design for Project Fair Go to replace that improvised "Shadow AI" use. Instead of an open-ended chatbot, we built a governed adaptation engine using structured class parameters, automated safety guardrails, and mandatory teacher review to generate classroom-ready materials in under five minutes.
The lesson plan isn't the problem.
Across Australia’s 531 special schools and developmental settings, more than one million students (over a quarter of all school enrolments) receive educational adjustments for disability or complex learning needs. For these students, standard curriculum resources don't work without heavy, individualised adaptation.
Teachers in these classrooms face a preparation workload that rarely gets acknowledged. For every topic, an educator must manually produce multiple differentiated versions: scaled across varying cognitive levels, adapted for sensory sensitivities, and mapped to augmentative and alternative communication (AAC) systems like Grid, PODD, and LAMP. Visual materials have to be sourced, figurative language stripped out, and physical sorting cards formatted, printed, laminated, and cut.
The Alannah & Madeline Foundation’s eSmart program offered high-quality digital safety lesson plans, but without setting-specific resources, turning a single lesson into usable classroom materials was taking teachers up to two hours.
The rise of "Shadow AI" in the classroom
To cope with the workload, exhausted educators were doing what anyone would do: turning to consumer AI tools like ChatGPT.
While raw LLMs offered quick text generation, they created an unvetted "Shadow AI" environment with serious structural flaws:
- Zero child safeguarding: Consumer models have no institutional guardrails to prevent harmful, confusing, or inappropriate outputs from reaching vulnerable students.
- Specialist curriculum blindness: Base models are trained on mainstream curriculum standards. They fail completely at developmental scales (such as Pre-Foundation levels A–D) and frequently inject idioms, metaphors, or Americanised references that derail neurodivergent learners.
- The prompt engineering burden: Expecting time-poor teachers to write flawless, multi-layered system prompts in an open text box is an unrealistic design expectation.
The real problem wasn't a lack of educational content. It was an adaptation gap that was pushing teachers toward high-risk, unregulated tools. AMF launched Project Fair Go to replace that improvised practice with a governed, purpose-built system.
The Research & Service Blueprint.
Reframing the brief from the classroom floor
Before touching any interface or prompt architecture, we ran in-person co-design workshops with specialist educators in Melbourne to map their daily planning routines. We asked six teachers to bring their real lesson plans and walk us step-by-step through their adaptation process.
That discovery revealed an immediate disconnect between the Foundation's original project brief and the reality on the ground:
- The lesson structure wasn’t the bottleneck: The Foundation originally assumed teachers needed better, more flexible lesson plan templates. In reality, teachers found the eSmart lesson structure sound.
- The time sink was multi-modal asset creation: The real drain was the manual labour required to make a lesson deliverable: rewriting abstract concepts into literal language, building visual sequence decks, and translating core vocabulary into symbols for augmentative communication devices (Grid, PODD, LAMP).
- Adaptation follows a strict, repeatable taxonomy: Teachers weren't improvising randomly. They systematically adapted materials against four clear variables: developmental level (such as Victorian Pre-Foundation levels A–D), communication system, sensory constraints, and high-interest engagement hooks.
Bridging user workflow to AI architecture via the Service Blueprint
To turn these qualitative insights into technical requirements, I developed a future-state Service Blueprint. In an Applied AI product, the blueprint serves as the critical bridge between UX and engineering, defining the relationship between educator actions, front-end states, backend LLM orchestration, and child safeguarding policies.
The blueprint established three core system principles:
- Structured parameters over prompt engineering: Instead of asking teachers to invent prompts, the UI would translate the educator's mental model into structured input parameters that feed deterministic prompt templates.
- Pre- and post-inference safety checkpoints: The system mapped explicit boundaries where student data had to be stripped before inference, and where generated outputs had to pass automated content evaluation before reaching the UI.
- Mandatory human-in-the-loop review: Full AI automation in a specialist classroom is unsafe. The blueprint explicitly decoupled content generation from classroom export, enforcing a mandatory review and edit stage where the educator retains complete authority over the final material.
- 01
Lesson Selection
Educator Goal & ActionsFind an AMF eSmart lesson to adapt; select base lesson & state/region.
Educator InputsBase lesson State / curriculum regionAI Processing LayerLesson structure parsing
Safety Controls (Automated AI Gates)Source grounding & hallucination checks
Primary JTBD OutcomeVerification — Ensure base lesson is safe, vetted, and policy-compliant.
- 02
Set Adaptation Parameters
Educator Goal & ActionsConfigure cohort needs; select parameters describing learner groups.
Educator InputsDevelopmental level Accessibility & sensory needs Learning supports (AAC)AI Processing LayerPedagogical adaptation engine
Safety Controls (Automated AI Gates)Pedagogical rule enforcement
Primary JTBD OutcomeScaling — Automate differentiation for the entire class.
- 03
Review & Refine
Educator Goal & ActionsReview proposed adaptations, adjust parameters or prompt, and regenerate.
Educator InputsEducator feedback Inline editsAI Processing LayerAdaptation explanations & editable outputs
Safety Controls (Automated AI Gates)Output safety shields & misuse filtering
Primary JTBD OutcomeConfidence — Educators review and approve adaptations before use.
- 04
Export
Educator Goal & ActionsSelect desired format (
HTML/PDF/PPTX) and download classroom resources.Educator InputsTarget resource formatAI Processing LayerStructured multi-format resource generation
Safety Controls (Automated AI Gates)AI disclosure & mandatory human approval
Primary JTBD OutcomeEfficiency & Independence — Ready-to-use classroom materials that enable self-learning.
"It's too hard to continually adapt everything. I wish people would develop resources for specialist schools rather than us having to take steps to continually adapt." Specialist Educator, Co-Design Workshop
Designing with guardrails.
An open chat box is the wrong interface for specialist education. Expecting an exhausted teacher to write complex prompts invites erratic outputs and risks off-curriculum hallucinations.
Instead, we built parameter scaffolding. Teachers select a verified base lesson from the eSmart library and configure their class context through structured controls to define parameters like developmental levels (including Pre-Foundation A–D), communication systems (Grid, PODD, LAMP), and sensory constraints. The interface translates these selections into governed system prompts behind the scenes.
To keep the system child-safe, we worked with Sitback and Microsoft engineers to build guardrails directly into the Azure pipeline:
- Pre-inference: Student data is automatically de-identified, and prompts are locked to AMF's verified pedagogy.
- Post-inference: Automated filters strip out figurative idioms, American spelling, and foreign references that confuse neurodivergent students, adhering to the National AI Centre's AI6 principles.
- Human-in-the-loop review: Teachers can edit generated text directly or use a scoped conversational assistant for micro-adjustments before exporting printable resources.
We designed this through the lens of the "curb-cut effect." Building an adaptation engine for the complex demands of specialist classrooms creates a resilient, governed foundation that can eventually scale to mainstream schools nationwide.
"This is infrastructure for fairness, not just an AI tool."
- Select Lesson: Choose a base lesson from the eSmart catalog.
- Set Parameters: Define IEP goals, AAC tools & sensory needs.
- Review & Refine: Inspect AI outputs with teacher oversight.
- Export: Download slides & printable materials.
The right problem, solved.
The discovery research and co-design workshops proved our initial assumption was incomplete. Educators didn't need more static lesson plans; they needed relief from the manual grind of adapting resources for diverse needs. That pivot directly shaped the MVP strategy, ensuring the design solved what teachers actually needed on the ground.
The co-design workshops confirmed that resource creation, not lesson structure, was the primary time sink. This finding redefined the project from static templates to an intelligent, parameter-driven generator, directly addressing the risks of educators relying on unvetted consumer tools like ChatGPT.
The future-state service blueprint and UI architecture provided the complete foundation for development. The resulting Azure-hosted MVP reduced preparation time from over two hours to under five minutes per lesson, with all five specialist educators in usability testing confirming immediate classroom readiness.
Rather than relying on generic outputs, every interface decision and pipeline checkpoint enforced child safeguarding and pedagogical integrity. Aligned with the National AI Centre's AI6 principles, the system recorded zero safety violations across all testing rounds while preserving 100% of the Foundation's core learning intentions through mandatory human-in-the-loop review.
Keep reading.
National One Stop Shop
Clinical trials in Australia run across eight different regulatory systems. I led the UX design of a Federal initiative to consolidate all of it into a single national platform.
Read the case study →Assisted School Travel Program
Staff coordinating transport for students with disabilities were working across multiple systems that didn't talk to each other, and the seven-second workflow that replaced two-and-a-half minutes of manual work.
Read the case study →