🧠 Routine AI

Built in 2 hours at Agents for Impact Hackathon · March 16, 2026 Powered by NVIDIA

Live Demo →

---

What It Does

Just rant about your day — Routine AI turns it into a clean, time-blocked schedule.

No forms, no dropdowns, no structure required. You dump whatever's on your plate and the app figures out the rest:

"I have an exam tomorrow and two homework assignments due at midnight, but I also want to do yoga at 6 and I have class at 2:30 and a meeting at 10:30 AM"

That's it. Hit the button and get back a full structured day.

---

Features

• 🎙️ Voice-first input — tap the mic, rant about your day, get a schedule back
• ⌨️ Text fallback — prefer typing? tap "or type" for a minimal inline textarea
• 📷 Scan mode — photograph a to-do list, supplement label, or anything printed; OCR reads it and builds a routine from it
• 🕐 Time-of-day context — morning / afternoon / night toggle so the schedule fits where you're at
• 📋 Animated schedule output — time-block cards with color-coded tags (focus, urgent, break, meeting, admin, personal)
• 📊 Summary stats — tasks, focus hours, and breaks at a glance
• 🗂️ History — past routines saved locally so you can reference what you built before

---

Tech Stack

Layer	Tool
Frontend	React + Vite
Deployment	Vercel
AI Scheduling	NVIDIA Nemotron (nvidia/llama-3.1-nemotron-ultra-253b-v1)
Speech-to-Text	NVIDIA ASR API
Image OCR	NVIDIA Infer API

All APIs are NVIDIA — called via https://integrate.api.nvidia.com/v1.

---

Getting Started

1. Clone the repo

git clone https://github.com/your-username/routine-ai.git
cd routine-ai

2. Install dependencies

npm install

3. Add your API key

Create a .env file in the root:

VITE_NVIDIA_API_KEY=your_nvidia_api_key_here

4. Run locally

npm run dev

---

How It Works

Voice Mode

1. Tap the mic → speak freely about your day
2. Audio sent as base64 to the NVIDIA ASR (speech-to-text) endpoint
3. Transcript appears below the mic as a floating caption
4. "Make my routine" sends it to Nemotron

Scan Mode

1. Tap the camera icon → file picker opens (camera on mobile)
2. Image sent as base64 to the NVIDIA OCR endpoint
3. Extracted text is passed to Nemotron with a scan-specific prompt
4. A checkmark preview replaces the camera icon once processed

Text Mode

Tap "or type" → inline text area expands → type your brain dump → submit

Schedule Output (all modes)

Nemotron returns structured JSON:

{
  "summary": { "tasks": 6, "focus_hours": 3, "free_blocks": 2 },
  "blocks": [
    {
      "time": "9:00 AM",
      "duration": "30 min",
      "title": "Meeting prep",
      "note": "Review notes beforehand",
      "tag": "focus"
    }
  ]
}

Cards animate in one by one with a staggered slide-up effect.

---

System Prompts

Brain dump:

"You are a calm, no-nonsense personal scheduler. Return ONLY raw JSON, no markdown. 6–10 blocks, always include a break and a meal."

Scan mode:

"You are a routine expert. Build a practical ordered routine from the scanned label text — skincare, supplements, meal prep, workout, cleaning, anything. Flag conflicts in notes."

---

Roadmap

• Voice as the default input (text as opt-in fallback)
• End-of-day reflection check-in
• Auto-prioritization
• Calendar integration with follow-up reminders
• Optional stress score / productivity rating

---

Hackathon

Built at Agents for Impact, hosted by NVIDIA · March 16, 2026
Nemotron requirement: ✅

---

License

MIT