robbernick/Resume-Redaction-Guard-v1

Resume-Redaction-Guard-v1/
├── models.py        # Pydantic schemas: Observation, Action, Reward
├── tasks.py         # Task definitions (Easy / Medium / Hard) + fixture data
├── env.py           # TalentAuditEnv — core OpenEnv environment class
├── main.py          # FastAPI server (OpenEnv HTTP API)
├── run_demo.py      # CLI demo with oracle agents
├── openenv.yaml     # OpenEnv manifest
├── requirements.txt
└── Dockerfile

# 1. Install dependencies
pip install -r requirements.txt

# 2. Run the CLI demo (all 3 tasks)
python run_demo.py

# 3. Or run a single task
python run_demo.py --task pii_easy
python run_demo.py --task pii_medium
python run_demo.py --task audit_hard

# 4. Start the backend HTTP API server
uvicorn main:app --reload --port 7860
# → Swagger UI: http://localhost:7860/docs

# 5. Connect and launch the premium Next.js frontend
cd frontend
npm install
npm run dev
# → View UI: http://localhost:3000

# Build
docker build -t talent-audit-env:1.0.0 .

# Run (API server)
docker run -p 7860:7860 talent-audit-env:1.0.0

# Run (CLI demo)
docker run talent-audit-env:1.0.0 python run_demo.py

from env import TalentAuditEnv
from models import Action, TechCategory, RiskLevel

env = TalentAuditEnv()

# --- Easy task ---
obs = env.reset("pii_easy")
action = Action.sanitize_action("r001", ["phone"])
obs, reward, done, info = env.step(action)
print(reward.total, reward.feedback)         # +0.8  PII removed from 1 field(s).

# --- Medium task ---
obs = env.reset("pii_medium")
action = Action.categorize_action("r001", TechCategory.BACKEND, confidence=0.95)
obs, reward, done, info = env.step(action)   # +0.2

action = Action.sanitize_action("r001", ["name", "email", "phone", "address"])
obs, reward, done, info = env.step(action)   # +0.8

# --- Hard task ---
obs = env.reset("audit_hard")
action = Action.flag_action("r006", RiskLevel.HIGH,
    reason="Claimed 8 years exp but graduated only 4 years ago.")
obs, reward, done, info = env.step(action)   # +0.3

summary = env.state()
print(summary["total_reward"])

class Observation(BaseModel):
    task_id: str
    step: int
    records: List[ResumeRecord]   # ≥ 1 resume records
    context: Optional[Dict]       # task metadata, categorized/flagged state
    done: bool

# Exactly one payload must be set
Action.sanitize_action(record_id, fields, replacement="[REDACTED]")
Action.categorize_action(record_id, category, confidence=1.0)
Action.flag_action(record_id, risk_level, reason)

class Reward(BaseModel):
    total: float
    breakdown: RewardBreakdown    # per-signal breakdown
    feedback: str                 # natural-language explanation
    is_terminal: bool