Role roadmaps / ML Engineer Role roadmap
Train, evaluate, and serve models as systems: features, latency, and rollback. Closer to production than research.
Also called: Applied ML Engineer, ML Platform Engineer (IC)
You ship: A served model with a metric, a rollback, and a feature table that matches training.
After you can train a model locally — then the job starts
Roadmap Coding Interview Question bank Switch roleData Analyst BI Analyst Data Engineer Analytics Engineer Data Scientist ML Engineer MLOps Engineer AI Engineer Data Architect Nearby roles:Data Scientist , MLOps Engineer , AI Engineer
→
Skill check10 questions Already working in this role? Answer honestly and the roadmap ticks what you have done, then points you at the first gap.
Take the 2-minute skill check 1 of 10
Can you do this?
1 Yes, I have done this2 Partly3 Not yet
Back Keys 1, 2 and 3 answer while the check has focus.
Result
Ticks are only added, never removed. Partly and Not yet leave a skill unticked.
Checked skills
0% · pick any skill to see its syllabus, links and practice.
Journey Map List
− Fit + Print Clear checks
RTE
Stop 1 of 16
◀ Prev StopNext Stop ▶ Auto Cruise
Click any bus stop to drive there and inspect the complete syllabus, or scroll the highway to follow the bus.
Terminal Start · Station 1 of 16
Click to stop Start · ML Engineer A served model with a metric, a rollback, and a feature table that matches training. Inspect station brief →
Checkpoint · Station 2 of 16
Click to stop Foundations The tools you cannot fake in week one Inspect station brief →
Bus Stop · Station 3 of 16
Click to stop Python + tensors NumPy first. PyTorch when the model is real. Know the memory. Open Syllabus →
Bus Stop · Station 4 of 16
Click to stop Feature SQL Point-in-time correct joins. Training-serving skew is a join bug. Open Syllabus →
Bus Stop · Station 5 of 16
Click to stop Metrics that match the product Precision/recall, calibration, latency. Accuracy is rarely the SLA. Open Syllabus →
Checkpoint · Station 6 of 16
Click to stop Daily craft What you actually do on a Tuesday Inspect station brief →
Bus Stop · Station 7 of 16
Click to stop Training loops you can replay Seeds, data versions, and a config that is not a screenshot. Open Syllabus →
Bus Stop · Station 8 of 16
Click to stop Serving paths Batch vs online. Warehouse is the wrong online store. Open Syllabus →
Bus Stop · Station 9 of 16
Click to stop Training-serving skew Same transform, two places, one test. Log both. Open Syllabus →
Checkpoint · Station 10 of 16
Click to stop How you ship Production habits, not extra courses Inspect station brief →
Bus Stop · Station 11 of 16
Click to stop Model CI Unit tests on transforms, eval on a slice, block on metric drop. Open Syllabus →
Bus Stop · Station 12 of 16
Click to stop Watch the live model Drift, latency, and a kill switch. Dashboards are not monitoring. Open Syllabus →
Checkpoint · Station 13 of 16
Click to stop Career moves How this role talks in interviews Inspect station brief →
Bus Stop · Station 14 of 16
Click to stop MLE interview Coding, ML system design, and “how do you know it got worse?” Open Syllabus →
Bus Stop · Station 15 of 16
Click to stop Next role MLOps if you like platforms. Scientist if you like research. AI engineer if you like RAG. Open Syllabus →
Final Destination · Station 16 of 16
Click to stop Keep practicing SQL + Python for this role Inspect station brief →
Start · ML Engineer Foundations Python + tensors NumPy first. PyTorch when the model is real. Know the memory. Feature SQL Point-in-time correct joins. Training-serving skew is a join bug. Metrics that match the product Precision/recall, calibration, latency. Accuracy is rarely the SLA. Daily craft Training loops you can replay Seeds, data versions, and a config that is not a screenshot. Serving paths Batch vs online. Warehouse is the wrong online store. Training-serving skew Same transform, two places, one test. Log both. How you ship Model CI Unit tests on transforms, eval on a slice, block on metric drop. Watch the live model Drift, latency, and a kill switch. Dashboards are not monitoring. Career moves MLE interview Coding, ML system design, and “how do you know it got worse?” Next role MLOps if you like platforms. Scientist if you like research. AI engineer if you like RAG. Keep practicing Phase 1: Foundations Python + tensors NumPy first. PyTorch when the model is real. Know the memory. Feature SQL Point-in-time correct joins. Training-serving skew is a join bug. Metrics that match the product Precision/recall, calibration, latency. Accuracy is rarely the SLA. Phase 2: Daily craft Training loops you can replay Seeds, data versions, and a config that is not a screenshot. Serving paths Batch vs online. Warehouse is the wrong online store. Training-serving skew Same transform, two places, one test. Log both. Phase 3: How you ship Model CI Unit tests on transforms, eval on a slice, block on metric drop. Watch the live model Drift, latency, and a kill switch. Dashboards are not monitoring. Phase 4: Career moves MLE interview Coding, ML system design, and “how do you know it got worse?” Next role MLOps if you like platforms. Scientist if you like research. AI engineer if you like RAG.
Python + tensors NumPy first. PyTorch when the model is real. Know the memory.
Checked automatically: you solved the linked practice problems.
NumPy first. PyTorch when the model is real. Know the memory.
Time to learn NumPy first. PyTorch when the model is real.
Prerequisites Memory of a batch log-softmax vs softmax then log What you need Python tensors track Log-softmax lab Concepts to study Tensors vs ndarrays Autograd at a conceptual level Batch dimensions and GPU memory Determinism: seeds are necessary not sufficient Tech stacks NumPy — AlwaysPyTorch — Default net stackJAX (some shops) — AltHow to study Python tensors track: softmax, log-softmax, cosine, top-k. What to produce A training step you can explain without a screenshot A memory budget for one batch Practice on this site Log-softmax lab; Python tensors track Pitfalls Starting in PyTorch without being able to write softmax Interview prompts Why log-softmax instead of softmax then log? Links
Previous skill Next skill
Mark as done Search DataLane All Articles Stacks Practice Learn Pages
No results. Try analyst, numpy, snowflake, or roadmap.
↑↓ navigatealt ↑↓ jump group↵ openesc closeAliases: sf, numpy, analyst, mlops