InvisibleHer

Inspiration

One of us tutors a Year 6 girl in a regional town. Last term she scored top of her class in physics. Last week she told us she wants to drop science in Year 9.

We asked her why.

She said — "it's not for girls like me."

Nothing in her marks said that. Something in her room did.

We sat with that sentence for a long time. Girls like me. Eleven years old and she'd already done the maths — added up every "alright guys", every textbook example where a man pushes a cart, every time a teacher's eyes slid past her hand to call on a boy who hadn't even raised his — and decided physics wasn't a place she was allowed to stand in.

She wasn't wrong about her room. She was wrong about herself.

So we asked a simple question: what if the leak in the STEM pipeline isn't about confidence or aptitude or "reach" — what if the rooms we've built are just quietly, politely, kindly telling girls they don't belong in them? Each signal is small. The damage is the volume. Four thousand hours of it before Year 10 subject selection.

You can't fix what you can't see. So we built the thing that sees it — and the thing that rewrites the room she walks into next.

This is for her. And for every girl-like-her who's already done the maths.

What it does

InvisibleHer does two things, on purpose.

1. The Empathy Simulator. A 60-second interactive classroom. Pick a card — Aanya (9, neurodivergent, regional NSW), Jess (9, Gomeroi, First Nations), Maya (14, Vietnamese-Australian), Leila (14, wheelchair user). You live a Year 4 maths class or a Year 9 physics class. The teacher talks. The classmates talk. A side panel quietly tallies the bias signals as they happen. By the end you've collected 17 of them in 60 seconds. That's the hook — we make you feel it before we sell you the fix.

2. BiasLens. Drop in any worksheet, slide, or paragraph. An AI scores it on a 0–100 Bias Footprint across six dimensions we wrote ourselves: gendered language, representation gap, stereotype reinforcement, default-male examples, missing intersectional voices, and tokenism. Then one click rewrites it inclusively — and the substitute names come from a hand-curated Mirror Library of 16 verified, living intersectional women in STEM. Ten of them are Australian (Michelle Simmons, Karlie Noon, Marita Cheng, Veena Sahajwalla and others). No hallucinations. No tokens. Real role models only.

   How we built it

   Frontend: React 19, Tailwind, framer-motion for the simulator and the Mirror profile cards. We went editorial — cream, ink, electric coral — instead of the usual purple-gradient SaaS look. The simulator is dark and cinematic. The product surface is bright and journalistic. They feel like two different worlds because they are. Backend: FastAPI + MongoDB. Six API endpoints: /api/scene, /api/scene/run, /api/samples, /api/scan, /api/rewrite, /api/mirrors. The AI bit: OpenAI's gpt-5.2 via the Emergent Universal LLM Key. Two prompts — one for scoring against our rubric, one for the inclusive rewrite. Both forced to return strict JSON. The IP bit (the part that's actually ours): the six-dimension Bias Rubric, the curated Mirror Library, and three pre-seeded demo samples with stable scores so we never depend on Wi-Fi at the venue. Process: we wrote the rubric first, on paper, before any code. That decision saved us — the LLM never got to define the framework. It just executes ours.

   Challenges we ran into

   The simulator state machine fought us for four hours. Our first version would advance through the dialogue beats, hit the choice card, and immediately skip to the closing screen — showing "you chose to stay quiet" even when nobody had clicked anything. Turned out we were updating two pieces of React state in the same tick. Rewrote the whole thing into explicit phases (intro → choice → outcome → closing → done) and it worked.

The LLM kept hallucinating women. Early versions of BiasLens would invent a "Dr. Sarah Chen, Australian astronomer" who didn't exist. We learned the hard way: you can't ask an LLM to be intersectional and accurate at the same time. So we built the Mirror Library by hand. Now the model can only pick from a list of 16 real, living, sourced human beings. No exits.

Picking the women was harder than we thought. We wrote a list of 30 names. Half were dead. A third were American. We wanted the first ones an Australian Year 8 student would see to be Australian and alive — Karlie Noon, Michelle Simmons, Veena Sahajwalla, Marita Cheng (who literally founded Robogals — the org running this hackathon). Took two of us a full afternoon to verify every claim against a real source.

Writing the simulator dialogue without it feeling preachy. We pulled phrases from real Australian classroom audits and from our own memories of Year 9. "Boys, see me at lunch about engineering club." "Good lad, that's the engineer brain right there." "May-ah, did your old man help you with the homework?" If it sounded performative, we cut it.

Accomplishments that we're proud of

We named a metric. Bias Footprint didn't exist this morning. By tomorrow morning a few hundred people will know it does. All eight intersectional communities from the brief are in the product — not as a checkbox in marketing copy, but in the simulator dialogue, the Mirror Library tags, and the rubric itself (Dimension 5 mathematically penalises tokenism). The whole thing works end-to-end on a real LLM in a browser. Score, rewrite, mirror suggestion, side-by-side diff. Live. Not a mockup. It's beautiful. No AI-slop purple gradients. Editorial typography (Fraunces + JetBrains Mono), heavy whitespace, asymmetric layouts. We argued for an hour about whether to keep the cream colour. We kept it. We didn't pretend to solve all 7 stages of the leaky pipeline. Our MVP covers Primary School and High School — the first 3 stages — and we labelled the other 4 honestly as roadmap. We think Completeness is judged on honesty about scope, not breadth.

What we learned

The leak isn't dramatic. It's a thousand small kindnesses delivered with a smile. "Girls don't usually like this, but give it a go anyway." The kindness is the camouflage. An LLM is an executor, not an oracle. Give it a rubric and a sourced library, or it will make things up. Designing for intersectionality changes the whole product. If we'd built this for "girls" generically, we would have built another mentorship app. Building it for a Gamilaraay astrophysicist's classroom and a wheelchair user told the lab isn't set up for her forced a totally different architecture. Silence is the loudest thing you can put in a 3-minute pitch. We learned the names of 16 women in STEM we should have already known. That one stings.

What's next for InvisibleHer

The Pivot Moment — a 3-minute interactive at the Year 10 subject-selection decision (the visible leak point), which auto-generates a briefing card for the school's career counsellor. One product, three influencers: student, counsellor, parent. A school dashboard — track Bias Footprint trend across a term. Anonymised teacher leaderboard. Weekly nudge emails. PDF and image upload — drop in a photo of a textbook page; OCR + scan in one go. A public methodology page — full rubric, sources for every Mirror, audit trail. We want teachers to be able to challenge our scoring. That's the point. Real teachers — we want to put this in front of 10 NSW Year 8 science teachers next term and watch what happens to their Bias Footprint over 12 weeks. Stages 4 to 7 — undergrad, entry, mid-career, senior. The leak doesn't stop at Year 12. Eventually, neither will we.

Built With

• fastapi
• mongodb
• radix
• react-19
• react-router-7
• tailwind-css
• uvicorn