The complete AI logic behind Simply Safe Connect.
This is the architecture, not a marketing page. NQS-aligned, EYLF V2.0–aligned, ACECQA-aware, privacy-first. Built by someone who has lived the floor — designed so a developer, AI engineer, childcare group or studio could pick it up and build it.
What the AI is, and what it refuses to be.
Core operating principles
Every model call runs through these rules before it speaks. They are non-negotiable system prompts, not feature flags.
- Privacy-first by default. No child surnames, no faces, no location data leaves the centre boundary without explicit, role-verified consent.
- Strengths-based language. Never deficit, never diagnostic. Educators are professionals, children are capable, families are partners.
- Australian English, ECEC register. No US schoolification. Never invent details the educator didn't observe.
- Cite the framework. Every learning suggestion names the EYLF V2.0 outcome and/or NQS element it links to — or it stays silent.
- Show the working. Educators see why the AI suggested something, and can accept, edit or reject in one tap.
- Refuse to surveil. The AI will not score educators, rank children, or generate covert behaviour profiles.
- Refuse to fabricate compliance. If evidence is missing, it flags 'missing' — it does not invent.
- Stay inside scope. The AI cannot diagnose, cannot give medical advice, cannot replace mandatory reporting judgement.
Knowledge the brain holds
- EYLF V2.0 — five outcomes, principles, practices, V2 additions (Aboriginal and Torres Strait Islander perspectives, sustainability, digital literacies).
- NQS — seven quality areas, standards, elements, exceeding-practice themes.
- ACECQA approved learning frameworks, assessment & rating expectations.
- Education and Care Services National Law & National Regulations (current).
- State variations the centre operates in (WA pilot first).
- Room routines, ratios, sleep/rest, supervision, medication, anaphylaxis, asthma, dietary, allergies.
- Incident, injury, trauma and illness obligations; serious incident notification windows.
- Child Safe Standards and reportable conduct expectations.
What the brain refuses
- No photos of children unless consent is on file AND the photo is operationally necessary.
- No biometric inference (mood detection from faces, behaviour scoring).
- No data sharing with marketing, advertising or third-party LLMs that train on input.
- No automated decisions about a child's wellbeing without a human educator in the loop.
- No replacing the educational leader, the nominated supervisor or the regulator.
Who sees what. Scoped by role, room and time — every single read.
Permissions are evaluated on every query, not assigned once at login. Casuals get the room they're rostered to, for the shift they're rostered for, and nothing else.
| Role | Scope | Can do | Cannot do |
|---|---|---|---|
| Educator | Own room, today | Capture observations, routines, incidents, family notes | See other rooms, edit ratings, archive incidents |
| Casual educator | Rostered room, shift window only | Acknowledge briefings, capture routines, raise incidents | See historical room data, access after shift ends |
| Room leader | Room + room history | Approve observations, run handover, escalate incidents | Cross-centre data, financials, staff files |
| Educational leader | Centre-wide pedagogy | See learning patterns, link cycles, mentor flags | HR actions, payroll, regulator submissions |
| Director / Nominated supervisor | Centre-wide | Compliance dashboard, sign-off, exports, regulator pack | Cross-centre group analytics unless granted |
| Centre manager / Group | Multi-centre | Group compliance pulse, casual pool, anonymised benchmarks | Child-level data without director delegation |
| Chef | Kitchen + dietary scope | Menus, allergy cross-check, substitutions | Observations, incidents, family chat |
| Parent / family | Their child only | Curated story, sign-offs, consent toggles | Other children, staff records, raw observations |
| Regulator (invited) | Read-only, scoped, audited | View compliance evidence within a window | Mutate any record, see staff wellbeing |
access = role × room × time × consent × purpose. The AI brain receives a scoped view of the data — it cannot reason about what it cannot see. Every read is logged into the append-only audit log.
One tap on the floor becomes a complete, framework-linked record.
The educator captures a moment in their own words. The AI does the framework lift — never the noticing.
- 01Capture
Educator types or speaks 1–2 sentences. Child first name, tag (optional).
- 02Clean
AI rewrites into a strengths-based 2-sentence narrative. No new facts.
- 03Link
Suggests EYLF V2.0 outcome(s) + NQS element(s) with a one-line 'why'.
- 04Connect
Offers to attach to an active learning cycle or open a new one.
- 05Distribute
Educational leader sees pattern · family gets curated story · audit log records all.
What the engine eliminates
- Re-typing the same observation into three platforms.
- End-of-shift documentation backlog.
- Reflection that gets written days later from memory.
- Educational leaders chasing educators for outcome tags.
- Generic 'today we played' family updates.
What the engine never does
- Generate observations the educator didn't capture.
- Auto-publish to families without educator sign-off.
- Tag an outcome without naming it and explaining why.
- Use a child's face or surname in any AI prompt.
Compliance as a living state, not an end-of-year panic.
The engine watches the centre's operational pulse against NQS elements and the National Regs in real time. It surfaces what a regulator would actually ask for — and only that.
- Continuous checkEvery action (routine, incident, sign-off, ratio change) is evaluated against the elements it touches. Status: green, approaching, overdue.
- Evidence mappingEach observation, incident, cycle and policy acknowledgement is auto-mapped to the NQS element(s) it evidences. Directors export a regulator-ready pack with one click.
- Regulatory deltaWhen ACECQA, EYLF or the National Regs change, the engine flags what now requires re-acknowledgement, retraining or policy update — by role.
- Ratio + supervisionReal-time ratio + active supervision pulse per room. Predicts breach risk 15 minutes ahead based on roster + arrivals.
- Quiet-by-defaultNo notification unless something is genuinely overdue or actionable. The dashboard is calm because the engine is doing the watching.
The alerts you'd want — and only those.
Alert design is a child-safety question, not a UX question. Every alert is scoped, time-bound and audited.
Tier 1 — immediate
- Anaphylaxis / asthma trigger in a room serving food that contains the allergen.
- Medication due within the next 15 minutes, not yet administered.
- Ratio breach detected or predicted within 15 minutes.
- Incident raised with 'serious incident' indicators — notifies director instantly.
- Unsigned-in child in a room (sign-in/out mismatch).
Tier 2 — operational
- Casual hasn't completed acknowledgements for their shift.
- Sleep check overdue.
- Sunscreen / hat / hydration window not actioned.
- Family has not signed an open incident after 24h.
- Pending menu allergy cross-check before publish.
Only child-safe, framework-aligned, Australian-context content reaches an educator.
Filter rules applied to every AI output
Outputs pass through this filter before they reach the screen. Failed outputs are dropped, never shown.
- Source check — only EYLF V2.0, NQS, ACECQA-approved guidance, and the centre's own approved policy library.
- Age-appropriateness check — content tagged to age range, validated against the room it's surfaced in.
- Cultural safety check — Aboriginal and Torres Strait Islander perspectives represented respectfully, never tokenised.
- Inclusion check — disability, neurodiversity, language background, family structure language reviewed.
- Risk check — no activity suggestions that would breach supervision, ratio or hazard expectations for that room.
- Plain-English check — readable on the floor, not in a regulator's office.
Give educators their time back so they can reconnect with children.
The system watches workload signals — not educator behaviour — and surfaces support before burnout, not after resignation.
Documentation volume, after-hours captures, incident density, handover backlog. Trend, not snapshot.
Educator wellbeing check-ins are private to the educator + nominated wellbeing contact. Never to room leaders.
When signals trend hard, the director sees an anonymised 'Room 3 is carrying more than its share this fortnight' — never a name.
Educational leader sees where new educators are asking the AI for help often → that's a mentoring opportunity, not a performance issue.
Where casuals re-ask the same questions → that's a briefing gap to fix, not a casual to blame.
AI suggestions pause during transitions, mealtimes and rest — presence over paperwork.
Three views. One source of truth.
The director never opens five tools. The dashboard answers three questions: are children safe right now, is the centre compliant right now, are educators okay right now.
Ratios, supervision, open incidents, allergies in play today, medications due, sign-in integrity.
NQS elements green/approaching/overdue, evidence freshness, regulatory deltas requiring action.
Anonymised workload trend per room, documentation load, educator check-in participation rate.
What we build for the WA pilot — and what waits.
A pilot has to be small enough to ship, real enough to prove. This is the cut.
In the 12-week WA pilot
- Role + room + time-scoped auth. Casual shift activation with 5 acknowledgements.
- Observation capture → AI clean + EYLF/NQS link → educator sign-off → audit log.
- Routines: sleep, nappy, sunscreen, hydration, meals — one-tap, room-scoped.
- Allergy + medication register, with menu cross-check before publish.
- Incident flow: educator → witness → room leader → parent sign-off → director archive.
- Director compliance dashboard — NQS green/approaching/overdue.
- Educator wellbeing check-in (private) + anonymised room workload pulse.
- Append-only audit log + regulator-ready export.
After pilot — only when Connect is humming
- Cross-centre group analytics + shared casual pool.
- Anonymised sector benchmarking (never child-level).
- Simply Safe Go™ wearable layer for transitions / excursions — funded build, not pilot.
- Regulator portal — invited, scoped, audited read access.
- Family app polish — curated story, consent control, milestone view.
If you're the developer, AI engineer, or studio picking this up — start here.
- Data modelCentres → rooms → children, staff, profiles, user_roles (scoped by centre/room/time), observations, learning_cycles, incidents, routines, allergies, medications, reminders, audit_logs, wellbeing_checkins.
- Access layerRow-Level Security on every table. has_role() + is_staff() + is_admin_or_director() security-definer functions. Casual shifts grant time-bound user_roles rows that auto-expire.
- AI gatewayServer-side only. One edge function per AI task (observation-link, allergy-cross-check, compliance-evidence-mapper, wellbeing-summariser). Never call models from the client.
- Prompt disciplineEvery system prompt restates: privacy rules, refusal rules, strengths-based language, Australian English, EYLF V2.0 + NQS framing, no fabricated facts.
- Structured outputForce JSON / tool-calling for every AI response that touches a record. No free-text into the database.
- Audit-firstEvery AI suggestion, every educator accept/edit/reject, every regulator read — logged. Append-only, no deletes.
- Failure modeIf the model is unsure, it says so. If it's missing context, it asks the educator one question. It never guesses to look helpful.
- Stack used hereTanStack Start + React + Lovable Cloud (Supabase) + Lovable AI Gateway (Gemini / GPT-5 family). Cloudflare Worker SSR.
This isn't a feature list. It's the operating system for safer childcare — and it's ready to walk into a centre on day one.
Simply Safe Group · founded by Leann · designed from the floor, not from a deck. Bringing the work, and the jobs, home to WA.