انتقل إلى المحتوى الرئيسي

AI Predictive Analytics

Project leadership wants the answer to one question: which projects are about to surprise us? The Predictive Analytics module answers it with four numpy-only predictive models that score every active project for risk.

Overview​

The module ships four predictors:

  • Delay predictor — likelihood of schedule slip ≥ 10%.
  • Cost overrun predictor — likelihood of cost ≥ 10% over baseline.
  • Quality risk predictor — likelihood of NCR spike next 30 days.
  • Safety risk predictor — likelihood of LTI next 30 days.

Each predictor is a gradient-boosted decision tree or random forest implemented in pure numpy — no Python ML stack required, deployable on any Python 3.13 runtime, including tenants with strict supply-chain policies.

Screenshot: Predictive dashboard — each project scored across four risk axes with traffic-light colors

Key concepts​

Risk score. Each predictor outputs a 0-100 score. Cutoffs (green / amber / red) are configurable per tenant.

Feature importance. Every prediction is explainable — the dashboard shows which input features contributed most to the score (e.g., "low SPI", "high NCR aging", "missing toolbox talks").

Refresh cadence. Models score every active project nightly. Manual refresh is available on the project detail.

Backtest. The model's historical accuracy is shown on the dashboard — useful for trust calibration.

Step-by-step: review the dashboard​

  1. Open Insights → Predictive analytics.
  2. The grid shows every active project with the four risk scores.
  3. Click any project to drill in: feature importance, historical scores, suggested actions.

Step-by-step: act on a prediction​

  1. Find the high-risk project.
  2. Review the top contributing features (e.g., "SPI 0.78", "NCR aging > 30 days").
  3. Use the suggested actions panel — typical actions are linked to the relevant module (e.g., "Open NCR register").
  4. Mark the prediction as Acknowledged with a free-text note. The acknowledgement is logged.

Common tasks​

  • Filter by score: Dashboard → Filter → Score > X.
  • Compare projects: Compare view shows scores side-by-side.
  • Export risk register: Export → PDF/Excel for project review meetings.
  • Tune cutoffs: Settings → Risk thresholds per tenant.
  • Configure auto-notify: any project flipping to red can auto-page the project director.

Troubleshooting​

  • "Project shows no scores." — Project has too little history (less than 4 weeks). Models need a baseline.
  • "All scores are amber." — Tenant cutoffs may be too tight. Tune at Settings → Risk thresholds.
  • "Feature importance shows nothing." — Model not yet trained for this tenant's data. Default cross-tenant model is used until tenant has 10+ closed projects.
  • "Backtest accuracy is low." — Models improve over time as more projects close. Until then, treat scores as advisory.

Limits​

  • Model retraining: monthly.
  • Max projects scored per nightly run: 10,000 per tenant.
  • Feature count: 60 features per project.

Permissions reference​

PermissionWho needs it
PREDICTIVE_VIEWPMs, Senior management
PREDICTIVE_ACKNOWLEDGEPMs
PREDICTIVE_THRESHOLD_EDITTenant admin

See also: Cost Estimation, AI Chatbot, Reports.