DataEngBytes 2023 - SYD-T1-02 - Mikiko Bazeley
summary
Argues companies should stop chasing Google scale and extend existing ML platforms to host LLM use cases using RAG and vector DBs rather than ripping everything out, showing concrete options: use off‑the‑shelf LLMs, fine‑tune, or train from scratch and choose serving patterns accordingly. Maps tooling to layers (BigQuery/Redis, Kubernetes, Airflow, experiment tracking, feature stores like Featureform) and cites industry signals (about 87.7% of orgs aiming to standardize on one AI platform). Warns against inventing a separate prompt engineer silo.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 3:36What is MLOps
- 6:05Core principles
- 11:21Why MLOps matters
- 15:54LLM Ops hype
- 18:40Data science workflow
- 23:16MLOps layers
- 25:25Solo vs team needs
- 27:50Integrating LLMs
- 33:05AI adoption strategy
- 38:24Prompt engineering & data quality
- 40:28Future of MLOps
related talks
12mDigital & Data/AI Strategy and Transformation
38mLLMs for Self-serve Analytics - Lessons Learned
29mRAG for Data Engineers: Why Your Metadata Is the Missing Piece
24mDemocratizing data access through LLM
24mSoftware 3.0 and the AI Engineering Landscape
23mThe State of GenAI & ML in the Cloud Native Ecosystem
40mTechniques to avoid common LLM pitfalls
32mDataEngBytes 2023 - PER-T1-11 - Mark Kitson