s-b3/LifeStack

<div align="center">

🪐 LifeStack

Autonomous Multi-Domain Conflict Resolution via Cascading RL

Built for Meta × HuggingFace PyTorch OpenEnv Hackathon 2026

  

**Live Demo** • **Technical Blog** • **Source Code**

</div>

🚀 The Vision

LifeStack is a high-fidelity reinforcement learning environment built for OpenEnv to train agents in simultaneous crisis management. Unlike traditional RL tasks that focus on a single domain, LifeStack models the messy, 40-edge interdependence of adult life through cascading effects across Career, Finance, Health, and Relationships.

✨ Core Research Innovations

• —🔗 Causal Cascades: 40-edge dependency graph based on Starcke & Brand (2012) where a $350 flight rebooking (Finance) ripples into stress (Wellbeing) and sleep loss (Health).
• —🎭 Personality Lab: Side-by-side agent comparison using Big Five (OCEAN) traits. Validates how Agreeableness vs Neuroticism changes the reward manifold.
• —🧠 Memory RAM: Retrieval-Augmented Moderation using ChromaDB. Shows a +116% improvement in strategy efficiency when recall is enabled.
• —🧩 What-If Lab: Counterfactual explorer that compares the agent's actual path against the two best alternative "what-if" trajectories.

🏗️ Hardened System Architecture

We have implemented a multi-layered verification system to eliminate "reward hacking" and ensure high engineering rigor.

🛡️ Anti-Hacking & Observability

• —Semantic Reasoning Audit: Every action requires a reasoning justification that is cross-verified for logical coherence by the reward orchestrator.
• —📼 Episode Replay: Full audit log of the last 5 episodes including metric impact grids and timestamped reasoning.
• —🌡️ Domain Risk Heatmap: Instant cognitive summary of 23 metrics across 6 life domains (Red=Crisis, Green=Stable).
• —🧪 Core Test Suite: 10 rigorous smoke and logic tests verify environment reset, causal propagation, and task solvability.

🗺️ Environment Map

graph TD
    subgraph "LifeStack Engine (v2.1)"
        Env["LifeStackEnv"]
        DG["Dependency Graph (40-Edges)"]
        RT["Route Manager"]
        RE["Reward Orchestrator (7-Signals)"]
    end

    subgraph "Observability Layer (Flask Portal)"
        CV["Cascade Visualizer"]
        WI["What-If Explorer"]
        Hist["Episode Historian"]
    end

    subgraph "AI Core"
        Agent["RL Agent / LLM"]
        Mem["ChromaDB RAG Memory"]
        Pers["Personality Engine (Big Five)"]
    end

    Agent -->|Action + Reasoning| Env
    Env -->|Cascades| DG
    DG -->|Feedback| Env
    Env -->|Verification| RT
    RT -->|Scoring| RE
    RE -->|Reward| Agent
    Agent <-->|Memory Store/Retrieval| Mem
    Observability <-->|Audit| Env

🛠️ Quickstart

1. Installation & Demo

git clone https://github.com/oki-dokii/LifeStack.git
cd LifeStack
pip install -r requirements.txt
python app_flask.py  # Production Portal → http://127.0.0.1:5000

2. Engineering Verification

# Run the full concrete logic test suite
python3 -m pytest tests/

3. Training Pipe (GRPO)

# Start 5-stage curriculum training with 800-word trajectory logs
python scripts/train_trl.py

📈 Performance & Results

RAG Memory Impact

Episodes were run back-to-back testing "Cold Start" vs "Memory-Aware" agents.

🏗️ Technical Deep Dive

• —Conflict Intake: Uses NLP-to-Conflict parsing; users can type natural language crises (e.g., "I just got fired...") and the system generates a personalized 23-metric disruption.
• —Observation Space: 26-dimensional state vector + domain-specific JSON metadata.
• —Reward signals: 7 non-overlapping components (Milestone, Completion, Outcome, Preservation, Replan, Efficiency, Reasoning) weighted iteratively for stability.

<div align="center">

Team BholeChature

Scaler School of Technology, Bangalore

<i>"LifeStack: Measuring the messy reality of human decision making."</i>

</div>