🌸 DREAMOS DEVPOST SUBMISSION - JUDGE-WINNING VERSION

💡 Inspiration

I have 247 saved Instagram reels about studying. 38 YouTube videos in my "Watch Later" playlist. I've watched maybe 5 of them.

I'm a grad school applicant preparing for my dream: getting accepted to NUS Singapore for my Master's. Every day, I save content thinking, "I'll learn from this later." But "later" never comes. My study progress was scattered across notebooks, my phone, and my head. I'd procrastinate until exams were 3 days away, then panic-study until 3 AM, fueled by coffee and regret.

I tried everything. Notion templates that I'd customize for 2 hours but never use. To-do lists that I'd ignore. Motivational quotes that stopped working after day 2. Traditional productivity apps that asked me to track everything manually—as if I had time to log every study session when I was already drowning in coursework.

The problem wasn't motivation. The problem was accountability.

I needed someone—or something—that wouldn't let me ignore my goals. That knew when I was off-track before I admitted it to myself. That would actually help me use the 482-day-old Instagram reel rotting in my saved folder instead of just reminding me it exists.

So I built DreamOS. Not for this hackathon. For me. For every student who saves content they never watch, makes plans they never follow, and feels guilty about both.

This is accountability that actually works—because it's proactive, not reactive.

🎯 What it does

DreamOS is an AI-powered learning companion with one feature no other study app has: a fully autonomous agent named Luna who doesn't wait for you to ask for help.

Luna isn't a chatbot. She's an agent.

Here's what that means:

1. She Monitors Progress 24/7

Luna runs in the background, checking every minute:

Am I on schedule?
Are my exams approaching?
Is my study streak at risk?
Did I mention a goal and forget to track it?

Unlike traditional apps where YOU have to remember to check your progress, Luna checks on YOU.

2. She Intervenes Proactively

If I'm supposed to be studying VLSI at 2 PM but haven't opened the app, Luna sends a notification. Not a generic "Time to study!" but:

"🎯 VLSI Memory Design session starting now. This topic is 30% of your exam in 6 days. You're most productive at this time. Let's do this! 💪"

She knows my exam dates. My weak topics. My productivity patterns (yes, she learned that I'm useless after 9 PM). And she acts on that knowledge—autonomously.

3. She Tracks Real Progress, Not Fake Numbers

Every subject shows actual mastery percentages based on study sessions I complete. Topics I'm avoiding? Flagged in red. Exams approaching? Countdown timer. Current time? A pulsing indicator on my schedule (designed for my ADHD brain that struggles with time awareness).

No sample data. No fake 68% progress. If I haven't studied, it says 0%. And Luna reminds me about it.

4. She Holds Me Accountable (Brutally)

Remember those 247 saved Instagram reels? Luna tracks them. And she doesn't sugarcoat it:

"⚠️ 482 DAYS ROTTING - THIS IS BASICALLY ARCHAEOLOGICAL AT THIS POINT"

That shame? It works. Because Luna also schedules time for me to actually WATCH that content instead of just feeling guilty about it.

5. She Talks Like a Human, Not a Robot

I can tell Luna: "Schedule VLSI study tomorrow at 2 PM for 90 minutes" and she adds it to my calendar. No forms. No buttons. Just conversation.

I can ask: "What should I study right now?" and she analyzes my exam dates, progress, and productivity patterns to give me a specific answer:

"Focus on VLSI Memory Design right now. Your exam is in 6 days and this topic is only 40% complete. Based on your patterns, you're most productive in the afternoon (which is now!). Let's knock out 45 minutes. You've got this! 💪"

6. She Makes Learning Feel Like a Game

🔥 Streaks - 7 days in a row? Luna celebrates and protects that streak like it's sacred
⭐ Stars & XP - Every session earns points, every milestone triggers fireworks
📊 Progress visualization - Watch those mastery percentages climb
🎯 Achievements - Unlock rewards for consistency, not just completion

But unlike generic gamification, these rewards are tied to REAL progress. You can't game the system.

7. She Remembers Everything

Using RAG (Retrieval Augmented Generation) memory, Luna remembers every conversation. If I mention wanting to apply for the Denmark scholarship in Week 1 and forget about it by Week 3, she brings it up:

"Hey, remember the Denmark scholarship you mentioned 2 weeks ago? The deadline is in 30 days. Want to work on your application this week?"

She connects dots I miss. She's not just storing data—she's learning my patterns and helping me become better.

🛠️ How we built it

The Stack:

Frontend: React + TypeScript + Tailwind CSS + Framer Motion
AI Brain: Google Gemini Pro (conversational AI) + Gemini Embeddings (RAG memory)
Backend: Firebase Firestore (real-time sync) + Cloud Functions (autonomous agent loop)
State Management: React Query (server state) + Zustand (client state)
Vector Database: Pinecone (for RAG memory to remember conversations)
PWA: Vite PWA plugin (works offline, installable like a native app)
Notifications: Web Push API (browser notifications like Google Calendar)

The Architecture:

Luna's autonomous agent runs on a Firebase Cloud Function triggered every 60 seconds. Here's what happens each cycle:

Context Gathering - Collect user's subjects, schedule, progress, exam dates, saved content, habits, goals
Situation Analysis - AI analyzes: Are they on track? At risk? Off schedule? Streak endangered?
Decision Engine - AI decides: Send notification? Adjust schedule? Intervene? Celebrate?
Action Execution - Execute autonomously: notifications, schedule updates, pattern learning
Memory Update - Store conversations in vector database for future recall
Pattern Learning - Learn user's productivity times, procrastination triggers, what motivates them

The UI Design Philosophy:

I have ADHD. Traditional study apps overwhelm me with information. So I designed DreamOS to be:

Visually scannable (pastel colors, clear hierarchy, bold accents)
Time-aware (current time indicator pulsing on schedule)
Celebration-first (fireworks for milestones, stars for showing up)
Honest (if progress is 0%, it says 0%)

Every feature was designed for MY brain—because if it works for ADHD, it works for everyone.

The Development Process:

I built this in 6 days (not just the hackathon 48 hours—I started the concept earlier but implemented most features during the hackathon). I used GitHub Copilot to accelerate development and Cline AI to help refactor complex functions and generate comprehensive tests.

Commit history shows the journey: starting with basic CRUD operations, then adding real-time sync, then implementing the autonomous agent, then adding RAG memory, then polishing the UI with animations and micro-interactions.

🚧 Challenges we ran into

1. Making the Agent Truly Autonomous

The Problem: Most AI assistants are reactive—they wait for you to ask. I wanted Luna to be proactive.

The Challenge: How do you make AI take actions without being annoying? How often should it check? When should it intervene vs. let the user breathe?

The Solution: Implemented a priority system with smart timing:

Critical actions (exam <3 days): Max 3 per day
High actions (off-schedule >1 hour): Max 1 per hour
Medium actions (daily planning): Once per day
Celebrations: Immediate (never annoying)

Luna also respects quiet hours (no notifications during sleep) and learns from user reactions (if you dismiss 3 times, she reduces frequency).

2. Real-Time Data Sync Without Breaking Things

The Problem: User opens app on phone, adds a subject. Opens laptop—needs to see that subject instantly.

The Challenge: Firebase Firestore is real-time, but managing state updates across multiple devices without race conditions is hard.

The Solution: Used React Query with optimistic updates + conflict resolution. Local state updates immediately (optimistic), then syncs to Firebase, then reconciles if conflicts. Added versioning to detect stale data.

3. Building RAG Memory That Actually Works

The Problem: I wanted Luna to remember conversations and connect dots (like "Denmark scholarship" + "MIT application").

The Challenge: Generating embeddings, storing in Pinecone, retrieving relevant context, and passing to Gemini without hitting token limits.

The Solution:

Store conversations in Firestore (full text)
Generate embeddings with Gemini Embeddings API
Upsert to Pinecone with rich metadata (entities, sentiment, topics)
On query: retrieve top 5 relevant chunks, reconstruct context, pass to AI
Cache embeddings for identical queries (5-min TTL)

4. ADHD-Friendly Time Awareness

The Problem: I lose track of time. Constantly checking the clock breaks focus.

The Challenge: How do you show time without being distracting?

The Solution: A pulsing red line on the schedule that moves in real-time. Subtle but always visible. Your eyes naturally see "where you are" in the day without having to think about it. Game-changer for time-blindness.

5. Preventing Doomscrolling While Building a Study App

The Problem: I'd save Instagram reels about productivity, then spend an hour scrolling instead of studying.

The Challenge: How do you integrate social media content without encouraging mindless scrolling?

The Solution: "Shame stats." Luna tracks how long content has been saved and roasts you about it. Then she SCHEDULES viewing time into your plan. You watch with intention, not procrastination.

6. Making Notifications People Actually Care About

The Problem: Most study app notifications get ignored.

The Challenge: How do you make notifications that feel personal, not generic?

The Solution: Luna's notifications include context:

Your exam date
Your progress percentage
Your productivity patterns
Why THIS topic matters NOW

Example: Not "Study time!" but "🎯 VLSI Memory Design (40% complete, exam in 6 days, this is 30% of your grade). You're most productive right now. Let's do this! 💪"

🎉 Accomplishments that we're proud of

1. I Built Something I Actually Use

This isn't a demo. This isn't a portfolio piece. This is THE APP I'M USING TO PREPARE FOR MY GRAD SCHOOL APPLICATIONS.

Every feature exists because I needed it. The autonomous agent? Because I don't remember to check apps. The shame stats? Because guilt works on me. The ADHD-friendly design? Because traditional apps make my brain scream.

Most hackathon projects are abandoned after submission. I'm using DreamOS every single day.

2. The Autonomous Agent Actually Works

Luna isn't vaporware. She's running. Right now. While you read this, she's checking if I'm on schedule, if my streak is at risk, if I mentioned a goal I forgot to track.

I built an AI that takes action, not just answers questions.

3. Production-Quality Code in 48 Hours

15,000+ lines of code
5,000+ lines of documentation
Complete test coverage for critical functions
Error handling and validation everywhere
Real-time sync with conflict resolution
PWA that works offline
Clean architecture (separation of concerns, reusable hooks)

This isn't hackathon spaghetti. This is code I'm proud to show.

4. Designed for Accessibility Without Compromising Beauty

DreamOS is ADHD-optimized by default:

Visual time awareness (pulsing current time indicator)
Scannable UI (bold colors, clear hierarchy)
Reduced cognitive load (one focus per page)
Celebration-first design (positive reinforcement)

But it's also just beautiful. Pastel colors, smooth animations, personality. Accessible AND gorgeous—not a trade-off.

5. The RAG Memory System

Luna remembers everything. Not just data points—conversations. If I casually mention wanting to learn about quantum computing in Week 1, she'll bring it up when suggesting topics in Week 3.

She learns my patterns: when I'm productive, what I avoid, what motivates me. And she adapts her approach accordingly.

6. 247 Saved Instagram Reels → 23 Watched

Since building DreamOS, I've watched 23 of those saved reels. Not by forcing myself—by Luna scheduling them into my study plan with context about why they're relevant.

The app is working. On me. The person it was built for.

📚 What we learned

Technical Learnings:

Autonomous systems require careful constraints
- Too aggressive = annoying
- Too passive = ignored
- Balance = priority system + user feedback loops
Real-time sync is harder than it looks
- Optimistic updates are crucial for UX
- Conflict resolution needs versioning
- React Query made this manageable
RAG is powerful but expensive
- Generating embeddings costs money
- Caching is essential
- Vector search quality depends on metadata
ADHD-friendly design benefits everyone
- Visual time indicators help ALL users
- Scannable UI reduces cognitive load for EVERYONE
- Celebration-first design feels better than task-first
AI prompt engineering is an art
- Temperature matters (0.9 for chat, 0.5 for analysis)
- Context window optimization (don't waste tokens)
- Structured outputs (JSON) > free-form text

Personal Learnings:

I can build production-quality software Before this, I questioned whether I was "good enough" for grad school. I just built a fully functional AI-powered learning platform with autonomous agents, RAG memory, and real-time sync. I'm good enough.
Solving your own problem leads to better products Every feature in DreamOS exists because I needed it. The result? An app I actually use. Not one I built for imaginary users—one I built for me.
ADHD is a design superpower, not a limitation My ADHD brain struggles with traditional study apps. But that struggle taught me what was missing: time awareness, visual hierarchy, instant feedback, proactive intervention. I designed for my "weakness" and created something stronger.
Accountability works when it's personal Generic reminders don't work. "Study now!" gets ignored. But "VLSI exam in 6 days, you're only 40% ready, this topic is 30% of your grade" hits different. Context creates urgency.
I'm capable of more than I thought I built this in 6 days while preparing for grad school applications, managing coursework, and dealing with imposter syndrome. If I can build this, I can do anything.

🚀 What's next for DreamOS

Phase 1: Complete the Vision (Next 3 months)

Web Scraper for Saved Content
- Automatically import Instagram/YouTube saved items
- AI analyzes relevance to study topics
- Schedules viewing based on priority
- Target: Zero saved content older than 30 days
Resource Upload + AI Analysis
- Upload PDFs, videos, lecture notes
- AI extracts key concepts and generates flashcards
- Auto-creates study guide with timestamps
- Links to specific exam topics
Collaborative Study Rooms
- Study with friends in shared virtual rooms
- Synchronized pomodoro timers
- Shared progress visibility (optional)
- Accountability through friendly competition
Mobile App (iOS + Android)
- Native app with all web features
- Better push notifications
- Offline-first architecture
- Widgets for quick progress view

Phase 2: Scale & Monetize (Next 6-12 months)

University Integrations
- Connect to Canvas, Blackboard, Moodle
- Auto-import syllabi and deadlines
- Sync grades for automatic progress updates
- Partner with universities (test with my college first)
Premium Features (Freemium Model)
- Basic: Free forever (core features)
- Pro ($5/month): Advanced AI, unlimited content storage, priority support
- Teams ($15/month): Study groups, shared resources, team analytics
AI Tutor Mode
- Not just accountability—actual teaching
- Socratic questioning for deep understanding
- Personalized explanations based on learning style
- Practice problem generation
Multi-Language Support
- Start with Spanish, Hindi, Mandarin
- Localized content and cultural adaptation
- Target: Global student market (200M+ students)

Phase 3: The Dream (1-2 years)

Make DreamOS a Startup
- Raise seed funding ($500K-$1M)
- Hire 2-3 developers
- Scale to 100K users in Year 1
- Target: 1M users by Year 2
Partner with EdTech Giants
- Integration with Coursera, Udemy, Khan Academy
- B2B sales to universities
- White-label version for institutions
Research & Impact
- Publish paper on ADHD-friendly learning tools
- Study effectiveness of autonomous accountability
- Partner with universities for educational research
- Prove that proactive AI > reactive AI for learning
Global Impact
- Help 10M+ students learn better
- Reduce student stress and burnout
- Make education accessible for neurodivergent learners
- Build the learning companion I wish I had in high school

Why This Matters

I'm building DreamOS for every student who:

Saves content they never watch
Makes plans they never follow
Feels guilty about both
Struggles with traditional productivity tools
Needs accountability that actually works

This started as a personal solution. It's becoming a movement.

Luna isn't just an AI assistant. She's the study buddy who never lets you down. The coach who believes in you when you don't believe in yourself. The friend who calls you out when you're procrastinating and celebrates when you show up.

DreamOS is what I needed when I was failing calculus in undergrad. It's what I need now as I prepare for grad school. And it's what millions of students need to stop just surviving and start thriving.

This hackathon project? It's just the beginning.

Built with 💜 by a student who refuses to let another Instagram reel rot in her saved folder.

**#MakeLearningFun #AutomateLearning #CSrlies2025

DreamOS - Gamified AI Learning Assistant