How to Become a ML Engineer
Deploys, scales, and monitors ML models in production. The bridge between research notebooks and real user impact.
1. Programming for ML
coreProduction-Grade Python
Beyond notebooks. Typing, testing, packaging, async. ML engineers ship code that's read more than written.
- docs free Python Docs →
- book paid Fluent Python →
SQL for Feature Engineering
Most training data starts as SQL against a warehouse.
- tutorial free Mode SQL Tutorial →
2. ML Fundamentals
coreClassical ML
Linear/logistic regression, trees, random forests, gradient boosting. Solves 80% of production ML problems better than deep learning.
- docs free scikit-learn User Guide →
- book free Introduction to Statistical Learning (ISLR) →
- docs free XGBoost Docs →
Model Evaluation
Metrics per problem type, cross-validation, calibration, cost-sensitive learning. Picking accuracy on an imbalanced dataset is the classic beginner error.
- docs free scikit-learn — Model Evaluation →
3. Deep Learning
corePyTorch
The production default. Tensors, autograd, DataLoader, distributed training.
- tutorial free PyTorch Tutorials →
- book free Dive into Deep Learning →
Transformers & Attention
The architecture behind LLMs, most SOTA vision models, and modern recommender systems.
- docs free Hugging Face Transformers →
- blog free The Illustrated Transformer →
4. LLMs & Generative AI
coreLLM Serving & Inference
vLLM, TensorRT-LLM, Ollama, llama.cpp. Quantization, batching, caching, throughput vs latency tradeoffs.
- docs free vLLM Docs →
- docs free Ollama Docs →
RAG & Vector DBs
Retrieval-augmented generation. Embeddings, vector search (pgvector, Weaviate, Pinecone, Qdrant), chunking strategies.
- tutorial free LangChain RAG Tutorial →
- repo free pgvector →
Fine-tuning
LoRA, QLoRA, full fine-tuning. When to fine-tune vs when to just prompt.
- docs free PEFT (Hugging Face) →
5. MLOps
coreExperiment Tracking
MLflow, Weights & Biases, Neptune. Reproducibility of runs is non-negotiable.
- docs free MLflow Docs →
- docs free Weights & Biases Docs →
Model Registry & Serving
Register versioned models, promote through stages, serve behind a stable API.
- docs free MLflow Model Registry →
- docs free BentoML Docs →
Training Pipelines
Airflow, Kubeflow, Metaflow, Prefect. Orchestrate feature engineering, training, evaluation, deployment.
- docs free Kubeflow Docs →
- docs free Metaflow Docs →
6. Feature Stores & Data
coreFeature Store Fundamentals
Feast, Tecton, Databricks Feature Store. Solve training/serving skew and feature reuse across models.
- docs free Feast Docs →
- blog free Tecton — What is a Feature Store →
7. Model Serving
coreServing Frameworks
TorchServe, Triton, KServe, Ray Serve, BentoML. Match to latency + throughput + hardware constraints.
- docs free NVIDIA Triton →
- docs free Ray Serve →
8. Monitoring & Observability
coreDrift & Model Monitoring
Data drift, concept drift, performance decay. Evidently, WhyLabs, Fiddler, Arize. Models silently degrade — monitor them.
- docs free Evidently AI Docs →
- blog free Arize AI ML Observability Course →
9. Distributed Training
recommendedDistributed PyTorch
DDP, FSDP, DeepSpeed. Scale beyond a single GPU/node.
- tutorial free PyTorch DDP Tutorial →
- docs free DeepSpeed Docs →
10. Cloud ML Platforms
recommendedAWS SageMaker / GCP Vertex AI / Azure ML
Managed ML platforms — trade lock-in for less infra work.
- docs free SageMaker Docs →
- docs free Vertex AI Docs →
11. Software Engineering Basics
coreTesting + CI/CD for ML
Unit test the code, integration test the pipeline, validate the model. GitHub Actions or GitLab CI for training + deploy.
- course free Made With ML — Testing →
12. Career & Reading
optionalApplied ML Blogs & Papers
Where ML engineers stay current.
- blog free Applying ML (Eugene Yan) →
- blog free Chip Huyen's blog →
- blog free Distill.pub →