Continuous Intelligence: Keeping your AI Application in Production
summary
Built a user‑facing used‑car price estimator with cross‑functional dev + data teams, three integrated pipelines (data -> ML training -> deployment), and automated retrain on git pushes. Uses Google Cloud Storage, DVC to point/version datasets, GoCD to orchestrate pipelines, Docker containers, and MLflow + ELK for metrics and monitoring. Model evaluation enforces an 80% accuracy gate so a switch to random forest or decision tree fails fast and prevents bad models reaching production.
AI-written summary · about
chapters (27)
- 0:00Introduction
- 3:40Continuous Delivery for Machine Learning
- 10:13Build Pipeline
- 13:01Challenges
- 15:01Reasons
- 17:45More types of change
- 19:10Pipelines
- 21:10Stages
- 25:39Technology stack
- 29:57Demo
- 30:40Delivery Pipelines
- 33:13Continuous Intelligence Workshop
- 35:53How it works
- 37:28Red pipeline
- 39:13DBC pull
- 41:13Redeploy
- 41:31Continuous Intelligent Cycle
- 44:23Questions
- 45:01How to package a model
- 46:43Bringing data scientists and developers closer
- 48:28Data versioning
- 50:26Evaluation stage
- 50:55Test stage
- 51:41Supervised learning
- 52:40Working together
- 55:01Continuous monitoring
- 57:18Monitoring
related talks
25mContinuous ML Applications in Production
29mData Science in Production
28mDataEngBytes 2023 - PER-T1-05 - Ellie Gholami
17mContinuous deployment for ML: The new software development lifecycle
35mHow to deploy machine learning models to production (frequently and safely)
16mOperationalizing AI in Finance: Real-Time Analytics
24mFrom Model to Production: Automating the ML Lifecycle with CI/CD
22mFrom experiment to production - a journey of a machine learning model