project.json

{
  "name": "Contextual Bandit Decision Simulator",
  "problem": "Teams need to validate decision policies offline before exposing users or systems to online reinforcement learning.",
  "domain": "reinforcement-learning",
  "architecture": "agent",
  "hugging_face_tasks": [
    "reinforcement-learning",
    "text-classification",
    "feature-extraction",
    "sentence-similarity"
  ],
  "recommended_stack": [
    "FastAPI for policy and feedback endpoints",
    "Vowpal Wabbit or River for online-learning baselines",
    "PostgreSQL for contexts, actions, propensities, and rewards",
    "Redis for low-latency policy serving",
    "MLflow for policy versioning",
    "OpenTelemetry for decisions and delayed rewards"
  ],
  "real_world_data_sources": [
    {
      "name": "Hacker News API",
      "url": "https://hacker-news.firebaseio.com/v0/topstories.json",
      "purpose": "Real content candidates for recommendation simulations"
    },
    {
      "name": "Open-Meteo API",
      "url": "https://api.open-meteo.com/v1/forecast?latitude=40.71&longitude=-74.01&current=temperature_2m",
      "purpose": "Real contextual features without authentication"
    }
  ],
  "job_description_skills": [
    "Contextual bandits and offline policy evaluation",
    "Propensity logging and counterfactual metrics",
    "Low-latency decision services",
    "Safe exploration and constrained actions",
    "Delayed-feedback pipelines and policy monitoring"
  ],
  "impact_targets": [
    "Beat a random policy by >= 20% average simulated reward",
    "Report IPS and doubly robust offline estimates",
    "Block 100% of disallowed actions before serving",
    "Serve policy decisions below 50 ms p95"
  ],
  "baseline_evaluation": {
    "test_examples": 4,
    "accuracy": 1,
    "synthetic_evaluation": true
  },
  "estimated_delivery": "8-12 weeks for one engineer",
  "generated_baseline_is_production_ready": false
}