Enhancing Quality Control Efficiency: A Dynamic Risk Threshold Approach
summary
Daily inflow of about 15,000 items/day meets a QC capacity of 200 so a system combines a per-item risk score (daily-retrained Random Forest with SHAP explanations) and a trolley-volume forecast to set a dynamic risk threshold in real time, served from Databricks endpoints and tracked with Airflow, Delta Unity Catalog, PySpark, MLflow and HyperOpt. Backtests show recall at K rising from ~2.5% (random) or ~4% (rules) to about 8.5–9%, with 10% of inspections left random to control bias.
AI-written summary · about
chapters (14)
- 0:00Royal Flora Holland intro
- 1:05Understanding Royal Flora Holland’s operations
- 2:54Quality control inefficiency
- 4:21Risk scoring supply items
- 6:27Dynamic risk threshold mechanics
- 9:41Adjusting risk thresholds in real time
- 12:55Trolley forecasting
- 14:46Tech stack and MLOps
- 15:07Model training and monitoring
- 18:18Real-time decisions
- 19:24Performance results
- 20:31Risk threshold demo
- 22:16Production deployment
- 23:27Q&A
related talks
15mExploring the math of thresholds
37mData Validation at Scale: Managing Data Quality in Complex Data Pipelines
16mStop Failures Before They Happen: Gradient Boosting
56mDecision Making in the Abscence of Ground Truth
43mPredictive Maintenance: From POC to Production With Spark
30mBetter data quality with ML & visualization
20mQuality Control to avoid GIGO in Deep Learning Models
43mAI-Driven Quality Engineering Practices