project.json

{
  "name": "Applied ML Support Router",
  "problem": "Support operations need reproducible routing models that expose confidence and defer uncertain cases.",
  "domain": "applied-machine-learning",
  "architecture": "classifier",
  "hugging_face_tasks": [
    "text-classification",
    "zero-shot-classification",
    "sentence-similarity",
    "summarization"
  ],
  "recommended_stack": [
    "FastAPI for prediction and feedback endpoints",
    "scikit-learn or LightGBM for production baselines",
    "Sentence Transformers for semantic fallback",
    "MLflow for experiments and model registry",
    "PostgreSQL for labels and human feedback",
    "Evidently-style drift and quality monitoring"
  ],
  "real_world_data_sources": [
    {
      "name": "GitHub Issues API",
      "url": "https://api.github.com/repos/pytorch/pytorch/issues?state=open&per_page=10",
      "purpose": "Real technical support-style issue text"
    },
    {
      "name": "Stack Exchange API",
      "url": "https://api.stackexchange.com/2.3/questions?pagesize=10&order=desc&sort=activity&site=stackoverflow",
      "purpose": "Real developer questions for weak-label experiments"
    }
  ],
  "job_description_skills": [
    "Applied NLP classification and confidence calibration",
    "Human feedback loops and active learning",
    "Model registry, drift monitoring, and retraining",
    "Feature, label, and evaluation pipeline design",
    "Business-aware automation and escalation metrics"
  ],
  "impact_targets": [
    "Reach macro F1 >= 0.85 on a reviewed support set",
    "Automate >= 60% of tickets at >= 0.90 precision",
    "Route low-confidence cases to human review",
    "Detect label and feature drift before SLA impact"
  ],
  "baseline_evaluation": {
    "test_examples": 4,
    "accuracy": 1,
    "synthetic_evaluation": true
  },
  "estimated_delivery": "8-12 weeks for one engineer",
  "generated_baseline_is_production_ready": false
}