sentinel-workforce-risk-models

Sentinel: Multimodal Workforce Risk Intelligence Models

This repository stores the production frozen model checkpoints, preprocessor parameters, PEFT/LoRA adapter weights, and calibrated late-fusion meta-classifier for Sentinel — an enterprise multimodal machine learning platform for workforce voluntary attrition and burnout risk prediction.

Main Application Repository: GitHub: ParminderSinghGithub/Sentinel

---

Model Components & Architecture

Sentinel combines structured organizational metrics and qualitative survey commentary using a calibrated multimodal late-fusion pipeline:

1. Tabular Branch (StructuredMLP):

   • Input Dimensions: 380 encoded features (24 continuous metrics + 356 one-hot categorical dimensions).
   • Topology: [380 -> 128 -> 64 -> 32 -> 1] with BatchNorm1d, ReLU, Dropout(0.20), and AdamW optimizer.
   • Target: Binary voluntary company exit (left_company).
   • Holdout Test ($N = 85,096$): ROC-AUC: 0.5755, PR-AUC: 0.3313, Log Loss: 0.5899, Brier Score: 0.2008, Recall at $\tau = 0.2469$: 84.70%.

2. Text Branch (DistilBERT + PEFT/LoRA):

   • Base Architecture: distilbert-base-uncased (66M parameters).
   • Adapter Config: Low-Rank Adaptation (LoRA, $r=16, \alpha=32$, target_modules=['q_lin', 'v_lin'], $\text{dropout}=0.05$).
   • Target: Psychological workplace distress and burnout indicator (high_burnout_risk).
   • Holdout Test ($N = 85,197$): ROC-AUC: 0.7363, PR-AUC: 0.7565, Log Loss: 0.6099, Brier Score: 0.2079, Recall at $\tau = 0.3530$: 86.46%.

3. Multimodal Late Fusion (MultimodalLateFusion):

   • Meta-Classifier: Calibrated Logistic Meta-Regression over unimodal log-odds: $$\text{logit}(P_{\text{exit}}) = 0.0094 + 1.0471 \cdot \text{logit}(P_{\text{structured}}) + 0.0272 \cdot \text{logit}(P_{\text{burnout}})$$
   • Operating Decision Threshold: $\tau^* = 0.2313$.
   • Aligned Dual Holdout ($N = 8,463$): ROC-AUC: 0.5719, PR-AUC: 0.3387, Recall: 86.60% (2,113 / 2,440 true departures captured).

---

Artifact Structure

├── structured_model/
│   └── best_checkpoint.pt              # PyTorch MLP weights + TabularPreprocessor state (251.5 KB)
├── text_transformer/
│   └── best_model/
│       ├── adapter_config.json         # LoRA hyperparameters (1.0 KB)
│       ├── adapter_model.safetensors   # Fine-tuned LoRA adapter tensors (3.55 MB)
│       ├── model_metadata.json         # Base model metadata (0.46 KB)
│       ├── tokenizer.json              # Subword vocabulary & tokenizer definition (711.7 KB)
│       └── tokenizer_config.json       # Tokenizer settings (0.31 KB)
├── fusion/
│   ├── fusion_model.joblib             # Calibrated LogisticRegression meta-model (1.05 KB)
│   └── evaluation_summary.json         # Verified holdout benchmarks & audit metadata (3.3 KB)
├── deployment_manifest.json            # Machine-readable SHA256 integrity manifest
└── README.md                           # Model card and technical specification

• Total Uncompressed Artifact Size: 4.31 MB across 8 runtime files.

---

Intended Use & Consumption

These artifacts are designed for offline inference consumption by the Sentinel runtime:

from workforce_risk.inference.predictor import WorkforceRiskPredictor
from workforce_risk.inference.schemas import EmployeeInput

# Initializes predictor using the local artifacts directory
# (or downloads snapshot from this repository if missing)
predictor = WorkforceRiskPredictor.from_artifacts("artifacts", device_str="cpu")

employee = EmployeeInput(
    employee_id="EMP-1001",
    department="Engineering",
    job_level="Senior",
    role="Senior Backend Engineer",
    tenure_months=28.0,
    salary=135000.0,
    satisfaction_score=0.78,
    performance_score=0.85,
    recent_feedback="Strong quarter overall. Team collaboration is productive."
)

result = predictor.predict_single(employee)
print(f"Risk Tier: {result.risk_tier} (Probability: {result.fused_risk_probability:.4f})")

---

Limitations & Ethical Considerations

• Scope: Designed strictly as an operational decision-support tool for HR leadership and team managers to proactively address workplace burnout and retention risks.
• Fairness & Privacy: Model inputs must exclude non-job-relevant demographic attributes. Individual employee predictions should not be used as automated employment termination criteria.