MediWatch

Inspiration

Truthfully this idea was something Ive always wanted to build but never got the chance to build it until this weekend. A while ago my grandmother fell at home one evening, and nobody knew for hours. By the time someone found her, what could have been a quick response turned into a serious medical emergency. That moment stuck with me.

1 in 4 seniors fall each year, and many of these incidents go undetected for critical minutes or even hours. I realized we needed technology that could be eyes where human eyes can't be, something that could actually save lives in those crucial moments.

What it does

MediWatch is an AI-powered patient monitoring system that watches hospital rooms and care facilities and personal homes through cameras, detecting medical emergencies in real-time.

Core Features:

• Real-time Emergency Detection: Continuously monitors video feeds to identify falls, choking incidents, seizures, and loss of consciousness

• Intelligent Voice Alerts: Announces emergencies out loud with specific details like "Fall detected in Room 203" using natural-sounding voice synthesis

• Smart Patient Triage: Ranks patients by risk level so nurses know who needs immediate attention and who's stable (using Woodwide)

• Sub-2-Second Response Time: From incident detection to alert, the entire pipeline completes in under two seconds critical for emergency response

• Email notification: sends an email to the primary person

🛠️ How we built it

Other Technologies

• Trae IDE we built it all in the Trae IDE, used builder mode and assigned multiple agents to different tasks

• Wood Wide AI for patient risk scoring and intelligent triage decisions

• Presage to detect heart rate from our video feed

• Gemini 3 Flash

• ElevenLabs for voice

Frontend & Infrastructure:

• Next.js
• TypeScript
• Tailwind CSS

Challenges we ran into

Performance Optimization: Getting video analysis fast enough without burning through API credits was brutal. We had to implement intelligent frame sampling (analyzing every 3rd frame instead of every frame) and optimize our prompts to reduce token usage.

False Positive Hell: Early versions thought everything was an emergency. A patient reaching for a glass of water? Fall detected. Someone rolling over in bed? Unconscious patient alert. We spent hours fine-tuning our Gemini prompts and adding context about normal patient movements.

Accomplishments that we're proud of

• Sub-2-Second Detection Pipeline: From incident occurrence to voiced alert in under 2 seconds—fast enough to make a difference
• Natural Voice Alerts: ElevenLabs integration produces alerts that sound calm and professional, reducing panic while maintaining urgency
• Hospital-Grade UI: Our interface looks professional enough that we'd feel confident demoing it to actual hospital administrators

Built With

• elevenlabs
• gemini
• nextjs
• traeide
• woodwide