DataEngBytes 2023 - PER-T1-05 - Ellie Gholami
summary
States that 85% of ML apps never reach production and recommends a concrete CI/CD pattern: develop code on sample data, promote code through Dev→Staging→Prod, then train on production data so restricted data never leaves Prod. Calls out tools: Delta Lake for data, MLflow for model and prompt tracking, LangChain/LlamaIndex or DSP for templating, GPUs and distributed fine‑tuning for cost, and using a vector DB as a separate service instead of retraining whole LLMs. Human feedback is required for continuous deployment.
AI-written summary · about
chapters (10)
- 0:00LLM apps and MLOps intro
- 3:19MLOps core principles
- 7:22Data-centric ML platform
- 11:23Code vs model lifecycle
- 15:40MLOps pipeline architecture
- 18:33LLM Ops vs MLOps
- 21:45Prompt engineering
- 24:01Packaging LLM models
- 25:27LLM compute optimization
- 26:36LLM deployment strategies
related talks
26mDataEngBytes 2023 - SYD-T2-03 - Akanksha Malik
41mCan Data & ML really learn from software engineering?
59mContinuous Intelligence: Keeping your AI Application in Production
41mInfrastructure for ML Applications
14mGetting Machine Learning applications into production using MLOps and CD4ML
29mML in Production on Databricks
39mDataEngBytes 2023 - BRI-T1-02 - Mikiko Bazeley
56mDevOps for Machine Learning