Building Sentinal: A Multi-Agent AI System to Break Doomscrolling

Google Cloud Gen AI Academy APAC Hackathon Project

Doomscrolling is something most of us are familiar with — you open Instagram or YouTube for “just 5 minutes,” and suddenly hours are gone. The real problem isn’t awareness — we know we’re wasting time. The problem is friction. There’s no immediate, personalized alternative that helps us shift behavior.

That’s exactly the problem we tried to solve.

🚨The Problem
Most productivity tools assume one thing:

You are already motivated. But what happens when you’re not?

When you’re stuck in passive consumption, you don’t want to open a task manager, create goals, or plan your day. You just keep scrolling.

We wanted to build something that intervenes at that exact moment.

💡 Our Idea: Sentinal
Sentinal is a multi-agent behavioral AI system that detects doomscrolling behavior and instantly converts it into a structured, actionable plan.

Instead of asking users to plan, it:

Understands behavior from natural language
Detects severity of distraction
Generates tasks, schedule, execution steps, and reflections
All in one interaction.

⚙️ How It Works
The system follows a structured pipeline:

1. User describes behavior
   “I spent 5 hours scrolling Instagram reels”

2. Doomscroll Detection Engine
   Rule-based scoring (no API)
   Classifies severity (Low → Critical)

3. Memory Injection
   Past sessions retrieved from SQLite
   Adds personalization context

4. Single LLM Call (Gemini 1.5 Flash)
   Generates outputs for 4 agents:
   Planner
   Scheduler
   Executor
   Reflector

5. Structured Output
   Validated using Pydantic
   Always returns clean JSON

6. Fallback System
   If LLM fails → rule-based output

7. MCP-style Tools Execution
   Tasks stored
   Notes saved
   Schedule created

8. Response to User
   Tasks
   Timetable
   First action step
   Behavioral insights

🧠 Architecture Overview
FastAPI backend (Python)
Rule-based detection engine
Gemini 1.5 Flash (single-call design)
SQLite for session memory
Docker + Google Cloud Run deployment

🔥 What Makes It Different?
Most tools:
Require manual input
Don’t detect behavior
Don’t intervene in real-time

Sentinal:
Detects behavioral drift
Works even without API (offline fallback)
Uses 4 agents in 1 LLM call (low latency)
Starts with zero friction first step

🚀 Key Features
Doomscroll Detection Engine
Planner Agent (task generation)
Scheduler Agent (time blocking)
Execution Agent (step-by-step guide)
Reflection Agent (behavior insights)
Session Memory
MCP-style tools (tasks, notes, calendar)
Offline fallback system

⚡ Challenges We Faced
Designing multi-agent logic within a single LLM call
Ensuring structured outputs using Pydantic
Handling LLM failures gracefully
Keeping latency low while maintaining quality
📚 Key Learnings
Single-call multi-agent design is highly efficient
Structured outputs (JSON mode) are critical for reliability
Behavioral AI is more impactful than traditional productivity apps
Fallback systems are essential for real-world deployment

🔗 Links
**GitHub: https://github.com/TulsiMundada/Sentinal-The-Autonomous-Life.git
Live Service: https://sentinal-service-6119332487701.asia-south1.run.app
👩‍💻 Team
Tulsi Mundada
Apoorva Barapatre

🌱 Final Thoughts
Sentinal is not just a productivity tool — it’s a step toward behavior-aware AI systems that understand users beyond commands.

Instead of waiting for motivation, it creates it.

And that’s where real impact begins.