Challenges in building operational AI
summary
Peltarion frames operational AI as far more than models: CSV parsing breaks pipelines, training often takes days or weeks while inference runs in a few hundred milliseconds, and simple demos (MNIST fits in 46 lines of code) mask production gaps. TensorFlow ships fast (eight major releases last year, five breaking changes), so reproducibility needs strict immutability, metadata and model lineage, horizontally scalable training, and cost controls like monitoring, preemptible instances and early stopping.
AI-written summary · about
chapters (37)
- 0:00Introduction
- 1:07AI will impact everything
- 1:24AI is a revolution
- 2:06AI disruption
- 2:24History of AI
- 4:42The importance of AI
- 5:04The magic of AI
- 6:26The Hello world of deep learning
- 7:17What is deep learning
- 8:40Developers vs developers
- 9:03CSV files
- 9:27Is it possible
- 10:31Gamma Knife
- 11:31Common problems
- 12:09What makes deep learning so versatile
- 12:42Deep learning is flexible
- 13:29Modeling is possible for nonexpert
- 15:29Deep learning takes time
- 16:05Is it ready for the real world
- 16:44Are we ready
- 17:06Operational phase
- 17:43Operational AI
- 18:49Machine Learning vs Software Engineering
- 21:10Data
- 22:05Existing tooling
- 23:26Dependencies
- 25:14Proof of concept
- 27:47Proof of concept workflow
- 28:40Operational system
- 31:28Immutability
- 32:11Deep Learning
- 33:29Scaling
- 34:24Data privacy
- 35:18Costs
- 36:13Restriction
- 37:42Where to start
- 40:59Framework
related talks
41mBeyond the Algorithm: What Makes Machine Learning Work?
16mOperationalizing AI in Finance: Real-Time Analytics
26mEngineering Excellence: Platforms and Architectures for Real-Time AI Analytics
25mAI In Production Key Challenges and Use Cases
42m3 Pillars of Open Source AI: Training, Inference, Agents
05mOvercoming challenges when implementing AI
36mApplied AI on the edge
16mAI: Common Misperceptions and Uncommon Perceptions