Delivering Delivery Time Prediction
summary
Built a delivery ETA service for Douglas using historical warehouse and carrier data to predict up to five business days, improving on a poor regression baseline by switching to XGBoost and a one‑vs‑all binary classifiers ensemble that emits per‑day confidence scores and uses probability cutoffs (or winner‑takes‑all) to produce a date range shown at checkout. Pipeline uses Minio, ModelDB, Argo pipelines and containers, deploys as a microservice in the client Kubernetes cluster via celldon, runs weekly manual retraining, and enters production in shadow mode with traffic/load checks.
AI-written summary · about
chapters (11)
- 0:00Delivery time prediction project
- 1:27Company and client overview
- 3:33Project motivation
- 5:19Warehouse process breakdown
- 7:05Shadow mode deployment
- 8:09Project pipeline and architecture
- 10:17ML approaches and experiments
- 13:06Training pipelines and tracking
- 14:32Front-end integration and results
- 15:34Key takeaways
- 18:47Q&A
related talks
12mUse ML to Predict Delivery Time and Improve Code Reliability
59mPredictability Check: Demand Forecasting
25mSupercharging Pipeline Efficiency with ML Performance Prediction
30mReal-Time ML services: From ideation to production
25mSupercharging Pipeline Efficiency with ML Performance Prediction
18mUsing Transformer Models to Predict Demand at Scale in Production
28mDemand Forecasting: Accuracy, Stability, Automation
58mThe path between developing and serving machine learning models.