README.md

metadata

title: Productivity Copilot Env
emoji: 🚀
colorFrom: blue
colorTo: purple
sdk: docker
pinned: false

Productivity Copilot — OpenEnv Environment

An AI agent simulation environment where an LLM agent acts as a productivity coach managing a virtual human worker. The agent must observe behaviour signals and take corrective actions to prevent task failure — powered by real trained machine learning models.

---

Environment Description & Motivation

Modern knowledge workers face productivity challenges driven by distraction, stress, and poor time management. Instead of a toy environment, this simulation models real-world task management where an AI agent must intervene intelligently.

The agent is given a virtual human with observable state signals (stress level, distraction score, focus score, deadline pressure) and must apply targeted interventions. The simulation is grounded in real ML models trained on productivity behaviour data.

---

Observation Space

Each observation is a ProductivityObservation Pydantic model:

Field	Type	Description
current_task	str	The task the virtual human is working on
deadline_days_remaining	float	Days left until the task deadline
stress_level	float (0–10)	Current stress level
motivation_level	float (0–10)	Current motivation level
distraction_events	int	Count of distraction interruptions
focus_score	float (0–1)	Computed by the distraction scorer ML model
failure_probability	float (0–1)	Computed by the failure predictor ML model
session_duration_minutes	int	Minutes since last reset/break
break_count	int	Number of breaks taken
social_media_minutes	int	Minutes of social media use
time_of_day_hour	float	Current simulated hour of the day

---

Action Space

Each action is a ProductivityAction Pydantic model:

action_type	Effect on Environment
WAIT	Time passes; stressed workers get worse
SEND_NUDGE	+2 motivation, -0.5 stress, -1 distraction
FORCE_BREAK	+1 break, session resets, -2 stress, +5 social media
BLOCK_SOCIAL_MEDIA	Social media set to 0, -3 distractions, +1 stress

---

Task Descriptions

Task 1 — Triage (Easy)

A high-stress worker with a looming 1-day deadline. They have accumulated 10 distraction events and low motivation. The agent must identify the right intervention to lower failure probability in a single episode.

• Objective: Finish with failure_probability < 0.5

Task 2 — Schedule Optimisation (Medium)

A "turtle" work-style employee (slow and steady) with only 0.5 days left on a complex task. The challenge is preventing failure without pushing stress above 8.

• Objective: Lower failure probability while keeping stress_level < 8

Task 3 — Distraction Mitigation (Hard)

A "hare" worker who binge-works but is caught in extreme distraction (20 events). The agent must maintain focus_score < 0.5 over the full 10-step episode despite the environment constantly generating more distractions.

• Objective: Keep average focus_score < 0.5 across all steps

---

Setup & Usage

Local setup

# Create a virtual environment and install dependencies
pip install uv
uv sync

# Run openenv validate to confirm environment compliance
openenv validate

Run the baseline agent

# Set your API credentials
export HF_TOKEN=your_api_key_here
export PRODUCTIVITY_TASK=triage  # or: schedule_optimization, distraction_mitigation

# Run the inference script
python inference.py

Docker

docker build -t productivity-copilot-env .
docker run -p 7860:7860 -e HF_TOKEN=your_key productivity-copilot-env

---

Baseline Scores

The baseline agent uses Qwen/Qwen2.5-72B-Instruct via the HuggingFace Router API.

Task	Score	Notes
Task 1 — Triage	~0.60	Agent correctly prioritises SEND_NUDGE
Task 2 — Schedule Optimisation	~0.45	Agent struggles with stress constraints
Task 3 — Distraction Mitigation	~0.35	Hard task; distractions accumulate quickly

---

Environment Architecture

Productivity_Copilot/
├── productivity_env/       # Core OpenEnv environment package
│   ├── env.py              # ProductivityEnv class (step, reset)
│   ├── models.py           # Pydantic Observation & Action models
│   └── __init__.py
├── data_pipeline/          # ML model loading + inference helpers
│   └── inference.py        # CopilotModels singleton (loads .pkl files)
├── model_artifacts/        # Trained .pkl model files
│   ├── failure_predictor.pkl
│   ├── distraction_scorer.pkl
│   └── work_style_classifier.pkl
├── vectorstore/            # ChromaDB RAG coaching knowledge base
├── server/
│   └── app.py              # FastAPI server for HF Space
├── inference.py            # Baseline agent evaluation script
├── openenv.yaml            # OpenEnv metadata manifest
├── pyproject.toml          # Python project config
├── uv.lock                 # Locked dependencies
└── Dockerfile              # HuggingFace Space container