ML plumbing
summary
Claims that most ML effort sits outside model code: 85% stay in POC and data scientists spend most time on cleaning while algorithms take 4% of work. Describes concrete plumbing to build: ingestion pipelines, single source of truth for cross‑departmental data, flexible event‑driven joins, automated cleansing and branching for experiments, production telemetry and thresholds, shadow mode and canary rollouts, user feedback capture and explainability, and automated retraining triggered by monitored drift.
AI-written summary · about
chapters (12)
- 0:00Introduction and speaker intro
- 1:05What is ML plumbing
- 3:57Why productionizing ML matters
- 7:06ML workflow and pipeline challenges
- 12:01Key capabilities for ML production
- 17:00Data abundance and silos
- 20:55Data correctness and bias
- 24:11Data flexibility and pipelines
- 29:05Model validity over time
- 31:13Building user trust in ML
- 34:46Summary and conclusion
- 36:32Q&A session
