Move safely. Feel confident. A gentle, camera-based AI coach that watches your form in real time and guides you through pregnancy-safe squats — out loud, hands-free, and personalized to your trimester.

→ Live Demo

Juno turns any laptop webcam into a prenatal personal trainer. It tracks your body with MediaPipe Pose, scores every rep against ACOG / Healthline–grounded medical rules tuned to your trimester, speaks coaching cues with ElevenLabs voice, and lets you ask questions out loud mid-workout — all wrapped in a calm, trauma-informed interface.

• Real-time form analysis — 33-point pose tracking at camera frame rate; live knee & hip angles, stance width, and knee-valgus detection drive a color-coded skeleton (green → adjust → stop).
• Trimester-aware coaching — depth windows, stance, and breathing cadence shift across 1st / 2nd / 3rd trimester. Relaxin-driven joint laxity means stricter knee-tracking thresholds — never push depth.
• Spoken cues, hands-free — "Keep your weight in your heels," "Lift tall through your chest," "Steady breathing." Delivered by ElevenLabs TTS with priority + cooldown so cues never talk over each other.
• Voice Q&A, any time — tap the mic between sets or mid-workout to ask about form, breathing, or setup. ElevenLabs Scribe transcribes → curated FAQ answers → spoken reply.
• Correct-form reference — a medically-reviewed Healthline squat demo sits beside the camera and its tip tracks whatever you're being coached on right now.
• Rep & set tracking — a standing → depth → standing state machine counts clean reps and runs a 3-set session with between-set breaks.
• High-impact safety — torso-velocity detection flags jumping / jerky motion as an additive warning that never blocks the normal form coaching.
• Pause / Resume — freeze the timer, analysis, and voice at any moment; resume cleanly with no cue backlog.
• Trauma-informed onboarding — pregnancy status, trimester, experience, fitness level, and doctor-clearance gate the experience; not cleared → a gentle breathing & education mode instead of live coaching.
• Safety first — an always-available ACOG stop-sign panel (dizziness, bleeding, contractions…) and a "this is not medical advice" posture throughout.
• Polished, calm UI — glassmorphism, mesh gradients, and an animated, accessible design system built on shadcn/ui + Radix.

flowchart LR
    A[Webcam] --> B[MediaPipe Pose<br/>33 landmarks]
    B --> C[Geometry<br/>knee/hip angles · stance · valgus · torso velocity]
    C --> D[Trimester × variation<br/>rule set]
    D --> E{Form cues}
    E -->|severity| F[Color-coded skeleton]
    E -->|priority + cooldown| G[ElevenLabs TTS cue]
    C --> H[Rep state machine<br/>standing → depth → standing]
    E --> I[Correct-form GIF + live feedback]
    H --> J[Set / session tracking → results]

Every frame, pose landmarks become geometric metrics, those metrics are checked against a rule set merged from your trimester and chosen squat variation, and the resulting cues drive the skeleton color, the spoken coaching, the reference panel, and the rep counter.

Server routes (TanStack Start handlers, deployed as Vercel functions) keep the ElevenLabs key server-side.

Coaching rules are grounded in ACOG ("Exercise During Pregnancy") and Healthline's medically-reviewed prenatal-squat guidance: keep weight in the heels, never force depth, watch for knee valgus (relaxin raises joint laxity), and stop at the first warning sign.

Depth & cadence by trimester (knee angle is hip–knee–ankle; smaller = deeper):

Form cues detected per frame: knees_over_toes · back_straight · knees_caving · too_deep · good_depth · chest_up · squeeze_glutes · remember_to_breathe · high_impact.

squat-form-analyzer/  (repo: Jade-Leong/teamFour)
├── src/
│   ├── routes/
│   │   ├── __root.tsx              # Document shell, head/meta, favicon, providers
│   │   ├── index.tsx              # Landing
│   │   ├── onboarding.tsx          # Trauma-informed intake (status/trimester/clearance)
│   │   ├── workout.$id.tsx         # Live coaching: pose, cues, pause, voice Q&A
│   │   ├── dashboard.tsx / progress.tsx / results.tsx / mood.tsx
│   │   ├── login.tsx / signup.tsx / profile.tsx / coach.tsx
│   │   └── api/
│   │       ├── tts.ts              # ElevenLabs text-to-speech proxy
│   │       └── stt.ts              # ElevenLabs Scribe speech-to-text proxy
│   ├── lib/
│   │   ├── pregnancyRules.ts       # Rule tables, pose math, cue detection, form GIFs
│   │   ├── voice.ts                # Client TTS/STT helpers + anti-overlap cooldowns
│   │   └── faq.ts                  # Curated Q&A matcher
│   ├── context/UserProfileContext.tsx
│   └── components/ui/              # shadcn/ui component library
├── vite.config.ts                  # TanStack Start + Nitro (Vercel preset)
└── vercel.json / .npmrc

Prerequisites: Node 18+ and an ElevenLabs API key (for voice).

# 1. Install (the lockfile needs legacy peer resolution)
npm install --legacy-peer-deps

# 2. Configure voice (create .env in the project root)
cat > .env <<'EOF'
ELEVENLABS_API_KEY=your_key_here
ELEVENLABS_VOICE_ID=EXAVITQu4vr4xnSDxMaL
EOF

# 3. Run the dev server
npm run dev          # → http://localhost:8080

Allow camera (and microphone, for voice Q&A) access when prompted, then Begin Workout.

vercel --prod

Set ELEVENLABS_API_KEY and ELEVENLABS_VOICE_ID in the Vercel project's environment variables. The build uses the Nitro Vercel preset (pinned in vite.config.ts) so SSR + /api/* routes deploy as serverless functions.

Juno is a fitness aid, not a substitute for medical advice. Always follow your clinician's recommendations, and stop immediately if you feel dizziness, pain, vaginal bleeding, shortness of breath, racing heartbeat, chest pain, fluid leaking, uterine contractions, or muscle cramps — then contact your healthcare provider.

Built for the VibeHack — Health Tech AI Hackathon. 💜 Repository: Jade-Leong/teamFour

Pose intelligence × pregnancy-safe medicine × a voice that actually coaches you.