Project Thumbnail

Inspiration

Last week, my five-year-old came home and said, "Mama, girls can't be scientists." She's five — and already being told what she can't be.

By age 6, research shows girls rate themselves as less talented than boys (Bian et al., 2017). Not because it's true — because that's what they've been told. And when your child looks up at you waiting for an answer, "that's not true" doesn't feel like enough.

I built EqualTales because every child deserves a better answer — not a lecture, but a story.

# What it does

EqualTales turns stereotypes into personalized, illustrated bedtime stories. A mother types what her child said (e.g., "Girls can't do math"), enters her child's name and age, and receives a 5-page watercolor-illustrated storybook featuring a real woman who proved that stereotype wrong.

The fictional protagonist — named after the child — discovers Marie Curie, Katherine Johnson, or Junko Tabei through a magical narrative arc: Belief → Question → Discovery → Inspiration → New Belief.

No signup. No login. Just stories that change beliefs.

# How we built it

5 MCP Tools via FastMCP for Goose integration:
- classify_stereotype — Claude Sonnet 4.5 categorizes the input
- match_real_woman — Local knowledge base of 50 verified women
- generate_story — Claude Opus 4.6 writes age-appropriate narratives
- verify_story — QA loop catches stereotype reinforcement
- generate_illustration — DALL-E 3 creates watercolor illustrations
Frontend: React 18 with SSE streaming, page-turn animations, typewriter text effects
Backend: Flask with parallel illustration generation (ThreadPoolExecutor)
Deployment: Vercel (frontend) + Render (backend)

# Challenges we ran into

AI amplifies stereotypes: Research shows AI-generated stories increase appearance-focused descriptions for girls by 55% (EMNLP 2025). We built a QA verification loop that checks every story before display.
Character consistency: DALL-E generates different-looking children across 5 pages. We solved this with detailed character description templates injected into every prompt.
75-second generation time: We parallelized QA + 5 illustrations using ThreadPoolExecutor, cutting total time from 150s to ~75s.
Goose integration vs. production reliability: We built a full MCP server that Goose can orchestrate, but use direct tool imports in production for demo stability.

# Accomplishments that we're proud of