L5 to L7: the AI upskilling path
The jump from L5 to L7 is not more code. It is scope: L5 owns a service, L6 owns a problem across services, L7 owns which problem is worth solving. What has changed is the subject matter - an increasing share of the systems that judgement is applied to now have a model in them.
For a decade the fastest route across the senior-to-staff gap was distributed systems. It still works, and it is still most of the job. What has changed is that the ambiguous, expensive, nobody-knows-the-answer decisions have partly moved into machine learning - and an engineer who cannot reason about a model is excluded from those decisions regardless of how strong the rest of their engineering is.
Nobody is promoted for knowing PyTorch. People are promoted for being right about what a system can and cannot do. The reason to learn this material is that it is now load-bearing for that judgement, not because the title asks for a framework.
The order below is the part that matters. Almost everyone starts at a transformer explainer, discovers it assumes linear algebra they last met at nineteen, and stops - then reads the restart as a talent problem rather than a sequencing one.
What to actually study
Real modules from the curriculum, in the order they make sense. Every video each one points at is free.
- DeploymentThe Language of SystemsThe five behaviours every production system exhibits, stated as maths before they are stated as advice.6 lessons
- Inference EngineeringThe Generation LoopUnderstand how an LLM produces text token by token --- from the forward pass and logits through sampling and stopping to the serving stack that puts it all in production.11 lessons
- Everything DataVector Databases & RAG PipelinesBuild retrieval systems that transform documents into embeddings, store them in vector databases, and serve them to LLM applications.7 lessons
- Everything DataData Versioning & ML ReproducibilityMake datasets, features, experiments, and model outputs reproducible across time, teams, and environments.7 lessons
Common questions
- How long does going from L5 to L7 actually take?
- The promotion is measured in years and is mostly about scope and evidence, not study. The knowledge gap this material closes is about three months of evenings - enough to reason about models honestly rather than defer on them.
- Do I need machine learning to make staff engineer?
- No. Plenty of staff and senior staff engineers work nowhere near a model. It matters when the ambiguous decisions on your surface area involve one, which is an increasing share of them, and being unable to size those proposals is what removes you from the room.
- Is this an interview preparation course?
- No. It teaches the material; it does not drill interview questions. The System Design track is the closest thing to interview-shaped, because system design interviews test the same reasoning.
No course promotes anyone. Levels are awarded by employers against their own ladder, on evidence collected over months. What material like this can do is close a specific knowledge gap that is currently in the way.
Every level