What to learn first
The fastest path is learning to decompose a workflow and connect the right data. Learn where judgment is needed, then evaluate whether the system made the business process better. Model-release news can wait.
- Start with workflows and the decisions users make.
- Design data access around retrieval, with permissions and source visibility handled up front.
- Use tool calling only where deterministic systems are better than text output.
- Evaluate on task success, with latency and cost tracked, and back it with regression checks and human review.
- Deploy with observability and rollback paths, under clear ownership.
AI engineering vs ML engineering
| Discipline | Primary question | Typical output |
|---|---|---|
| ML engineering | Can we train, serve, and monitor a model? | Models, pipelines, feature stores, model services. |
| AI engineering | Can we make an AI workflow useful and reliable? | Agents, tools, evals, interfaces, workflow automation. |
| Forward deployed engineering | Can we make it work inside a specific customer context? | Integrated systems, playbooks, production deployments. |
Glossary
- AI engineering
- The discipline of turning AI capabilities into reliable product and workflow systems.
- Agent
- An AI system that can use tools, follow a goal, and make bounded decisions across steps.
- Evaluation
- A repeatable test of whether an AI system completes the task correctly, and keeps doing so safely.
- Retrieval
- The process of giving a model relevant external context from documents, databases, or APIs.
- Human in the loop
- A control pattern where people approve, correct, or review AI output before high-impact actions.