CoolFace
Apppublic

saisatwik1234567/OpenENV

sourceHugging Faceupdated 6mo agoView on Hugging Face
0likes
App README

📧 OpenEnv: Email Triage & Response — Peak Edition

![Score](/) ![Tasks](/) ![Docker](/) ![Python](/) ![Tests](/)

Live Demo: https://saisatwik1234567-email-triage-and-response-environment.hf.space
bash
# Run in 30 seconds
docker build -t openenv-peak . && docker run -p 7860:7860 openenv-peak

🏆 Verified Scores

TaskDifficultyScoreThresholdResult
Task 1: Email Triage🟢 Easy1.000≥ 75%✅ PASS
Task 2: Response Drafting🟡 Medium0.889≥ 78%✅ PASS
Task 3: Inbox Zero Sprint🔴 Hard0.891≥ 68%✅ PASS
Average0.927✅ ALL PASS

🧠 Novel Technical Design

1. Synonym-Aware Keyword Grading

Unlike naive keyword matching, our grader uses synonym groups: refund ≡ reimburse ≡ credit ≡ money back — any word in the group = full credit.

2. Kendall Tau Prioritization Scoring

Mathematically rigorous rank-correlation metric: concordant_pairs / (concordant + discordant) Measures whether the agent processed urgent emails before low-priority ones.

3. Dense Shaped Rewards

Signal at EVERY step, not just episode end: reward = (grader_score - 0.5) * 0.8 with critical email penalty of -0.25.

🏅 What Makes This Different

FeatureThis ProjectTypical Submission
Email dataset50 curated real scenarios10-20 synthetic emails
GradingSynonym-aware NLUExact string match
PrioritizationKendall tau metricSimple ordering
Reward signalDense shaped ±1.0Sparse binary 0/1
DashboardLive real-time monitoringNone
Tasks3 progressive Easy→Hard1-2 basic tasks
API endpoints12 endpoints3-4 endpoints

📊 Dataset: 50 Enterprise Emails

urgent   ████████████  12 (24%) — P0 outages, GDPR, legal deadlines
normal   ███████████████ 15 (30%) — complaints, approvals, decisions
external ██████ 6 (12%) — partners, press, conferences
internal ███████ 7 (14%) — all-hands, IT, HR
low      ██████ 6 (12%) — newsletters, reminders
spam     ████ 4  (8%)  — phishing, fake domains

⚡ Quick Start

python
import httpx

BASE = "https://saisatwik1234567-email-triage-and-response-environment.hf.space"

# Reset
obs = httpx.post(f"{BASE}/reset?task_id=task_1").json()["observation"]

# Act
result = httpx.post(f"{BASE}/step?task_id=task_1", json={
    "action_type": "escalate",
    "email_id": "e001",
    "category": "urgent",
    "priority_score": 1.0,
    "escalate_to": "backend-oncall"
}).json()
print(f"Reward: {result['reward']}")  # +0.500

# Grade
grade = httpx.post(f"{BASE}/grader?task_id=task_1").json()
print(f"Score: {grade['episode_score']} | Grade: {grade['grade_letter']}")

🔌 API Endpoints (12 Total)

MethodEndpointDescription
POST/resetReset environment, returns Observation
POST/stepExecute action → (obs, reward, done, info)
GET/stateFull environment state
GET/tasksAll tasks with action schema
POST/graderGrade episode — score + letter grade + breakdown
GET/metricsLive episode stats
GET/leaderboardEpisode score leaderboard
POST/leaderboard/submitSubmit your agent's score
GET/infoFull environment spec
POST/baselineRun baseline agent (needs OPENAI_API_KEY)
GET/healthSystem status
GET/Live monitoring dashboard

🧪 Test Results

17/17 tests passing
pytest tests/ -v → 17 passed in 2.3s

📄 License

MIT — Open source, free to use for research and evaluation.