divAIne/the-busy-module-xgboost

Busy Detector — XGBoost Inference Endpoint

Model

Binary classifier predicting busy probability from 26 features (17 voice + 9 text).

Input

{
  "inputs": {
    "audio_features": {
      "v1_snr": 15.0,
      "v2_noise_traffic": 0.8,
      "v2_noise_office": 0.1,
      "v2_noise_crowd": 0.05,
      "v2_noise_wind": 0.05,
      "v2_noise_clean": 0.0,
      "v3_speech_rate": 3.5,
      "v4_pitch_mean": 150.0,
      "v5_pitch_std": 25.0,
      "v6_energy_mean": 0.1,
      "v7_energy_std": 0.05,
      "v8_pause_ratio": 0.3,
      "v9_avg_pause_dur": 0.8,
      "v10_mid_pause_cnt": 5,
      "v11_emotion_stress": 0.4,
      "v12_emotion_energy": 0.3,
      "v13_emotion_valence": 0.6
    },
    "text_features": {
      "t1_explicit_busy": 0.0,
      "t2_avg_resp_len": 8.5,
      "t3_short_ratio": 0.2,
      "t4_cognitive_load": 0.05,
      "t5_time_pressure": 0.0,
      "t6_deflection": 0.0,
      "t7_sentiment": 0.5,
      "t8_coherence": 0.8,
      "t9_latency": 1.2
    }
  }
}

Output

{
  "busy_score": 0.32,
  "confidence": 0.65,
  "recommendation": "CHECK_IN",
  "ml_probability": 0.28,
  "evidence_details": ["ML Baseline (-0.5)"]
}

Scoring Logic

Uses Evidence Accumulation (log-odds):

• Explicit busy intent: +6.0 weight
• Traffic noise: +3.0
• Emotion stress: +2.5
• ML model probability: ×0.5 log-odds factor

Final score = sigmoid(total evidence)

Thresholds

• Score < 0.3 → CONTINUE (person is free)
• 0.3 ≤ Score < 0.7 → CHECK_IN (uncertain)
• Score ≥ 0.7 → EXIT (person is busy)