Dopamine AI

A context-aware AI digital health assistant that helps college students fight doomscrolling through intelligent, graded interventions.

The Problem

College students lose hours daily to addictive apps. Existing screen time tools use rigid, one-size-fits-all limits that get ignored. They don't understand context — scrolling TikTok for 20 minutes on a Saturday morning is fine; doing it at 2 AM before an exam is not.

Our Solution

Dopamine AI is a context-aware digital health guardian. It considers what you're doing, when you're doing it, and what's coming up — then decides whether to leave you alone, nudge you, or block the app entirely.

Sense-Think-Act Pipeline

📡 Sense          →        🧠 Think           →        ⚡ Act
Collect context            AI risk assessment           Graded intervention
(app, time, calendar,      (LLM evaluates 0-3          (notification, popup,
 deadlines, duration)       using decision matrix)       or app block)

Level	Action	When
0	Pass	Using Google Docs during the day — keep going.
1	Gentle Notification	30 min on YouTube, no urgent deadlines.
2	Vibrate + Popup	Scrolling Instagram 15 min before lecture.
3	App Block	2 AM, exam in 6 hours, 90 min on TikTok.

---

App Walkthrough

1. Sign In

Users sign in with Google, which also grants access to their Google Calendar for context-aware scheduling.

2. Background Monitoring

Once set up, Dopamine AI runs in the background. When the user opens a distracting app, the Sense layer collects context:

• Current app and how long they've been on it
• Next calendar event and countdown
• Upcoming deadlines from Canvas
• Time of day — weekday night vs. weekend afternoon

3. Intelligent Interventions

Based on context, the AI decides the right level of intervention:

Level 1 — Gentle Nudge:

Level 2 — Popup Warning:

Level 3 — App Blocked:

---

Demo: Debug Console

We built a Streamlit-based debug console to demonstrate and test the full pipeline without needing a phone.

Running the Demo

conda activate aiHacks
streamlit run debug_ui.py

Tab 1: Sense-Think-Act Pipeline

The main demo tab. Use sliders and inputs to simulate any scenario, then watch the AI evaluate it in real time.

Step 1 — Set the context (Sense):

Use the left panel to configure:

• Which app the user is on (TikTok, Valorant, Google Docs, etc.)
• How long they've been using it
• Next calendar event and deadline
• Current time and whether it's a weekend

Or click a preset scenario button to load a typical situation instantly.

Step 2 — Run the pipeline:

Click "Run Pipeline". The AI evaluates the context and returns:

• Risk Level (0-3)
• Message to the user (in Tony's tough-love style)
• Reasoning for why it chose that level

Step 3 — Compare scenarios:

Change the parameters and run again. For example:

• Move "Continuous Minutes" from 20 → 90: watch the level go up
• Switch app from TikTok → Google Docs: watch the level drop
• Set hour to 2 AM + close deadline: level jumps to 3

Tab 2: Free Chat

Chat directly with Tony (the AI persona) to test conversational responses.

Tab 3: Prompt Lab

Edit the system prompt that drives the Think layer, and A/B test two different prompts on the same scenario side-by-side.

---

Architecture

┌─────────────────────────────────────┐
│         Flutter App (iOS/Android)   │
│  ┌──────────┐  ┌─────────────────┐  │
│  │ Google    │  │ Screen Time     │  │
│  │ Sign-In   │  │ Monitoring      │  │
│  └────┬─────┘  └───────┬─────────┘  │
│       │                │             │
│       ▼                ▼             │
│  ┌─────────────────────────────┐    │
│  │  Firebase Cloud Messaging   │◄───┤── receives interventions
│  └─────────────────────────────┘    │
└──────────────┬──────────────────────┘
               │ HTTPS (Cloud Function call)
               ▼
┌──────────────────────────────────────┐
│    Firebase Cloud Functions (Python) │
│                                      │
│    sense.py  →  think.py  →  act.py  │
│    (context)    (LLM eval)   (push)  │
│                                      │
│    wingman.py (pipeline orchestrator) │
└──────────────────────────────────────┘
               │
               ▼
┌──────────────────────────────────────┐
│    Oracle GenAI (LLM Inference)      │
│    Llama 3.3-70b / Cohere Command R+ │
└──────────────────────────────────────┘

Tech Stack

Layer	Technology
Frontend	Flutter (iOS/Android)
Backend	Firebase Cloud Functions (Python)
AI Model	Oracle GenAI (Llama 3.3-70b, Cohere Command R+)
Push Notifications	Firebase Cloud Messaging
Auth & Calendar	Google Sign-In + Calendar API
Debug UI	Streamlit

Project Structure

aiHacks/
├── .env              # API keys (Oracle GenAI)
├── wingman.py        # Cloud Functions entry — orchestrates the pipeline
├── sense.py          # Sense layer — context gathering & mock data
├── think.py          # Think layer — LLM risk evaluation with decision matrix
├── act.py            # Act layer — graded intervention execution
└── debug_ui.py       # Streamlit debug console for demo & testing