One of six curricula in the AI Career Curriculum
ML Engineering
You build and train the model: you decide what it learns.
9 role tracks and 82 modules, free and MIT-licensed.
- AI WorkforceWork alongside the models
- AI StartupTurn it into a company
- AI GovernanceGovern, secure, and assure AI
- AI EngineeringBuild with the models
- ML EngineeringBuild and train the models
- AI InfrastructureRun the platforms models live on
Your relationship to the model
To you the model is an artifact you shape: data, architecture, training objective, fine-tuning, and evaluation against the behavior you want. You change what the model knows and how it behaves, and you are the one who can say why it got better or worse.
This org covers ML engineering — building and training the models: data pipelines, classical and deep learning, fine-tuning and alignment, evaluation, distributed training, and the senior-to-principal work of setting ML architecture and strategy.
What you own
- Training data: sourcing, curation, labeling, and quality
- Training and fine-tuning: objectives, recipes, and experiments
- Model-level evaluation: benchmarks, regressions, and comparisons
- Model artifacts: versions, model cards, and lineage
- Choosing between training, fine-tuning, distilling, or using a vendor model as-is
What you don't own: The serving fleet, the product features built on top, or the final call on whether a use is acceptable. You hand off a model with evidence of what it can and can't do.
The tracks
Ordered by level. Each one opens with Module 00, which maps the role's place relative to the model before the technical work starts.
Foundation (L20)
ML foundations, data for ML, classical & deep learning, experiment tracking, deployment, drift monitoring
Specialist & Senior (L30–L35)
Data sourcing & licensing, curation & dedup, labeling ops, synthetic data, contamination checks, dataset versioning
SFT data, PEFT (LoRA/QLoRA), preference optimization, safety tuning, distributed fine-tuning, quantization
Benchmarks, LLM eval harnesses, LLM-as-judge, human eval, agent & safety evals, production eval
- NLP EngineerL30
Tokenization, classical NLP, classification, extraction, QA, summarization, machine translation
ML systems architecture, production LLM integration, experimentation at scale, ML SLOs, technical leadership
Distributed training, data pipelines at scale, cluster orchestration, checkpointing, elastic training, throughput
Staff & Principal (L40+)
Multi-team ML architecture, foundation-model training programs, ML platform strategy, capacity & TCO
Org & industry ML strategy, north-star ML architecture, influence without authority, the technical bar
Who you work with
No curriculum owns the model end to end. These are the handoffs with each of the others.
AI Workforce
Work alongside the models
- You hand them
- Models that are better at real work, plus honest documentation of what they're bad at
- They hand you
- Concrete examples of where model quality falls short on real tasks, the raw material for better evals and training data
AI Startup
Turn it into a company
- You hand them
- An honest read on what's technically feasible, how long it takes, and what it costs to own a model versus renting one
- They hand you
- The problems worth owning a model for, and the business case for the data and compute it takes
AI Governance
Govern, secure, and assure AI
- You hand them
- Data provenance, model cards, training-data documentation, and evaluation results they can audit
- They hand you
- Data-use rules, documentation standards, and the risk-based evaluation bar a model must clear
AI Engineering
Build with the models
- You hand them
- Models, fine-tunes, and embeddings with documented strengths, failure modes, and benchmark results on relevant tasks
- They hand you
- Real failure cases and task data from production, which tell them where a fine-tune or a different model would actually help
AI Infrastructure
Run the platforms models live on
- You hand them
- Model artifacts with known resource profiles (memory, throughput, precision) and the training workloads they need to schedule
- They hand you
- Training and fine-tuning clusters, experiment and artifact storage, and a path to production for the models they produce
What you'll work with
- Languages & frameworks
- Python, PyTorch, JAX
- Distributed training
- FSDP, DeepSpeed, Megatron, NCCL, CUDA, Triton
- Fine-tuning & alignment
- SFT, PEFT (LoRA/QLoRA), preference optimization, quantization
- Orchestration
- Ray, Kubernetes, Slurm
- Experiment tracking
- MLflow, Weights & Biases
- Evaluation
- Benchmark harnesses, LLM-as-judge, human evaluation, red-team evals
How the tracks work
Learn
Each track is a sequence of modules: lecture chapters, hands-on exercises, quizzes, and projects. Start with Module 00 and work forward.
Check your work
The paired solutions repo has a reference walkthrough for every exercise. Attempt each one first; the gap between your answer and the reference is where the learning happens.
Built in the open
The content is written with AI assistance and under ongoing human review, so it can contain mistakes. Report anything wrong through the repo's Issues; fixes ship continuously.