Learning path
AI/ML Engineer learning path: everything to study, topic by topic
Comprehensive AI/ML engineering assessment covering data and feature engineering, model development from classical ML to deep learning, LLM and generative AI systems, production ML serving, MLOps infrastructure, and responsible AI practices.
Short answer
To learn AI/ML Engineer, work through 7 modules and 35 topics in the order they build on each other. It starts with Data & Feature Engineering, then moves on to Model Development and LLM & Generative AI. Use the map as a checklist, or measure your level and study only what is missing.
What you will study
Topics are grouped by module. Open one to see its objectives and key concepts.
1 · Data & Feature Engineering5 topics
Data Collection & Preprocessing
Designing data collection pipelines, handling missing values, outlier detection, normalization, encoding categorical variables, and building reproducible preprocessing flows
Objectives
- Design preprocessing pipelines that handle common data quality issues
- Choose appropriate encoding and normalization strategies for different data types
Key concepts
- missing value imputation
- outlier detection
- normalization and standardization
- categorical encoding
- data quality validation
- ETL pipelines
Feature Engineering Techniques
Creating informative features from raw data, polynomial features, interaction terms, binning, target encoding, time-series feature extraction, and automated feature engineering
Objectives
- Create features that improve model predictive power
- Avoid data leakage during feature engineering
Key concepts
- polynomial features
- interaction terms
- binning strategies
- target encoding
- time-series features
- automated feature engineering
Feature Stores
Centralized feature management with Feast, Tecton, or SageMaker Feature Store, online vs offline stores, feature consistency, and point-in-time correctness
Objectives
- Design feature store architectures for training-serving consistency
- Implement point-in-time correct feature retrieval
Key concepts
- online vs offline stores
- point-in-time correctness
- feature consistency
- Feast
- Tecton
- feature serving latency
Data Versioning (DVC/LakeFS)
Versioning datasets and ML artifacts with DVC, LakeFS, or Delta Lake, reproducible experiments, data lineage tracking, and storage backends
Objectives
- Implement data versioning for reproducible ML experiments
- Design data lineage tracking across the ML lifecycle
Key concepts
- DVC pipelines
- LakeFS branching
- Delta Lake versioning
- data lineage
- reproducibility
- storage backends
Data Augmentation & Synthetic Data
Image augmentation, text augmentation, SMOTE for imbalanced data, GANs for synthetic data, diffusion-based generation, and evaluating synthetic data quality
Objectives
- Design augmentation strategies that improve model generalization
- Evaluate synthetic data quality and detect mode collapse
Key concepts
- image augmentation
- text augmentation
- SMOTE
- GAN-generated data
- diffusion models
- synthetic data validation
2 · Model Development5 topics
Supervised & Unsupervised Learning
Classification, regression, clustering, dimensionality reduction, model selection, cross-validation, bias-variance trade-off, and ensemble methods
Objectives
- Select appropriate algorithms for classification, regression, and clustering tasks
- Apply cross-validation and explain the bias-variance trade-off
Key concepts
- classification vs regression
- clustering algorithms
- cross-validation
- bias-variance trade-off
- ensemble methods
- model selection
Deep Learning Architectures (CNN/RNN/Transformer)
Convolutional neural networks, recurrent architectures, attention mechanisms, transformer architecture, architecture selection for different modalities
Objectives
- Choose appropriate deep learning architectures for vision, NLP, and sequence tasks
- Explain the transformer architecture and its advantages over RNNs
Key concepts
- CNN architectures
- RNN/LSTM/GRU
- self-attention
- transformer encoder-decoder
- positional encoding
- architecture selection
Transfer Learning & Fine-Tuning
Pre-trained model adaptation, feature extraction vs fine-tuning strategies, domain adaptation, learning rate scheduling, and catastrophic forgetting prevention
Objectives
- Design fine-tuning strategies that balance adaptation and knowledge retention
- Choose between feature extraction and full fine-tuning based on dataset characteristics
Key concepts
- feature extraction
- full fine-tuning
- domain adaptation
- learning rate scheduling
- catastrophic forgetting
- layer freezing
Hyperparameter Optimization
Grid search, random search, Bayesian optimization, Optuna, early stopping, learning rate finders, and multi-fidelity optimization strategies
Objectives
- Implement efficient hyperparameter search strategies
- Apply Bayesian optimization for expensive model training
Key concepts
- grid search
- random search
- Bayesian optimization
- Optuna
- early stopping
- multi-fidelity optimization
Experiment Tracking (MLflow/W&B)
Logging experiments with MLflow and Weights & Biases, metric tracking, artifact management, model registry, experiment comparison, and collaborative ML workflows
Objectives
- Set up experiment tracking for reproducible ML workflows
- Use model registries for versioning and deployment promotion
Key concepts
- MLflow tracking
- Weights & Biases
- metric logging
- artifact management
- model registry
- experiment comparison
3 · LLM & Generative AI5 topics
LLM API Integration
Integrating OpenAI, Anthropic, and open-source LLM APIs, handling rate limits, streaming responses, error handling, token management, and provider abstraction
Objectives
- Integrate LLM APIs with proper error handling and rate limit management
- Design provider-agnostic LLM abstractions
Key concepts
- API authentication
- rate limiting
- streaming responses
- token counting
- provider abstraction
- error handling
Prompt Engineering for Production
Production-grade prompt design, prompt templating, version control for prompts, systematic prompt testing, prompt optimization, and A/B testing prompts at scale
Objectives
- Design production prompt systems with versioning and testing
- Implement defenses against prompt injection attacks
Key concepts
- prompt templating
- prompt version control
- systematic testing
- prompt optimization
- A/B testing prompts
- prompt injection defense
RAG Architecture Design
Retrieval-augmented generation patterns, vector database selection, chunking strategies, hybrid search, reranking pipelines, and evaluating RAG quality
Objectives
- Design RAG architectures with appropriate retrieval and ranking strategies
- Evaluate and improve RAG system quality using standard metrics
Key concepts
- vector databases
- chunking strategies
- hybrid search
- reranking
- RAG evaluation
- context window management
Fine-Tuning & PEFT Methods
Full fine-tuning vs parameter-efficient methods (LoRA, QLoRA, prefix tuning), dataset preparation for fine-tuning, evaluation, and deployment of fine-tuned models
Objectives
- Choose between full fine-tuning and PEFT methods based on resource constraints
- Prepare high-quality datasets and evaluate fine-tuned model performance
Key concepts
- LoRA
- QLoRA
- prefix tuning
- adapter layers
- dataset preparation
- fine-tuning evaluation
Multi-Modal Systems
Vision-language models, audio-text systems, cross-modal embeddings, multi-modal RAG, and building systems that reason across modalities
Objectives
- Design multi-modal systems that combine vision, language, and audio
- Implement cross-modal retrieval and reasoning pipelines
Key concepts
- vision-language models
- CLIP embeddings
- multi-modal RAG
- cross-modal alignment
- audio-text models
- modality fusion
4 · ML Production Systems5 topics
Model Serving (TorchServe/Triton/vLLM)
Deploying models with TorchServe, Triton Inference Server, and vLLM, batching strategies, model optimization, and serving infrastructure design
Objectives
- Deploy models using production serving frameworks
- Optimize inference throughput with batching and model compilation
Key concepts
- TorchServe
- Triton Inference Server
- vLLM
- dynamic batching
- model optimization
- ONNX export
Model Monitoring & Drift Detection
Detecting data drift, concept drift, and model performance degradation, statistical tests for drift, monitoring dashboards, and automated retraining triggers
Objectives
- Implement drift detection pipelines with statistical tests
- Design automated retraining triggers based on monitoring signals
Key concepts
- data drift
- concept drift
- PSI
- KS test
- performance monitoring
- retraining triggers
A/B Testing for Models
Designing model experiments, canary deployments, shadow mode testing, statistical significance for model comparisons, and multi-armed bandits
Objectives
- Design A/B testing frameworks for model comparison in production
- Apply statistical methods to determine model superiority
Key concepts
- canary deployments
- shadow mode
- statistical significance
- multi-armed bandits
- interleaving
- guardrail metrics
CI/CD for ML Pipelines
Continuous integration for ML code and data, automated testing of models, deployment automation, rollback strategies, and ML-specific CI/CD patterns
Objectives
- Build CI/CD pipelines that validate data, code, and model quality
- Implement safe deployment and rollback strategies for ML models
Key concepts
- ML pipeline testing
- data validation in CI
- model validation gates
- deployment automation
- rollback strategies
- GitHub Actions for ML
Cost & Latency Optimization
Model quantization, pruning, distillation, caching strategies, GPU instance selection, spot instances, and inference cost optimization
Objectives
- Apply model compression techniques to reduce inference cost
- Design cost-optimal inference infrastructure with auto-scaling
Key concepts
- quantization (INT8/FP16)
- pruning
- knowledge distillation
- inference caching
- spot instances
- auto-scaling
5 · MLOps & Infrastructure5 topics
Containerized Training (Docker/K8s)
Dockerizing ML training workloads, Kubernetes for ML, custom training images, GPU scheduling, and reproducible training environments
Objectives
- Build containerized ML training pipelines with GPU support
- Design Kubernetes-based ML training infrastructure
Key concepts
- Docker for ML
- Kubernetes operators
- GPU scheduling
- custom training images
- reproducible environments
- KubeFlow
GPU/TPU Resource Management
GPU memory optimization, multi-GPU training, TPU configuration, mixed-precision training, and efficient hardware utilization
Objectives
- Optimize GPU memory usage for large model training
- Configure mixed-precision training for performance gains
Key concepts
- GPU memory optimization
- mixed-precision training
- gradient accumulation
- multi-GPU data parallelism
- TPU pods
- CUDA optimization
Distributed Training
Data parallelism, model parallelism, pipeline parallelism, DeepSpeed, FSDP, communication backends, and scaling training across nodes
Objectives
- Choose appropriate parallelism strategies for different model sizes
- Configure DeepSpeed or FSDP for distributed training
Key concepts
- data parallelism
- model parallelism
- pipeline parallelism
- DeepSpeed
- FSDP
- NCCL
Feature Pipeline Orchestration
Orchestrating feature computation with Airflow, Prefect, or Dagster, scheduling, dependency management, and feature freshness guarantees
Objectives
- Design feature computation pipelines with proper orchestration
- Ensure feature freshness and reliability in production
Key concepts
- Airflow DAGs
- Prefect flows
- Dagster assets
- scheduling
- dependency management
- feature freshness
ML Platform Design
Designing internal ML platforms, self-service ML infrastructure, platform APIs, multi-tenancy, governance, and balancing flexibility with standardization
Objectives
- Design ML platforms that balance flexibility with standardization
- Implement multi-tenant governance for shared ML infrastructure
Key concepts
- platform architecture
- self-service ML
- multi-tenancy
- governance
- abstraction layers
- developer experience
6 · Responsible AI5 topics
Bias Detection & Mitigation
Identifying bias in training data and model predictions, fairness metrics, pre-processing, in-processing, and post-processing debiasing techniques
Objectives
- Apply fairness metrics to detect bias in ML models
- Implement debiasing techniques at different stages of the ML pipeline
Key concepts
- fairness metrics
- demographic parity
- equalized odds
- pre-processing debiasing
- in-processing techniques
- post-processing calibration
Model Explainability (SHAP/LIME)
Local and global model explanations, SHAP values, LIME, attention visualization, feature importance, and communicating model decisions to stakeholders
Objectives
- Generate and interpret model explanations using SHAP and LIME
- Communicate model decisions to non-technical stakeholders
Key concepts
- SHAP values
- LIME
- feature importance
- attention visualization
- global vs local explanations
- stakeholder communication
AI Safety & Guardrails
Input and output validation, content filtering, safety classifiers, hallucination detection, guardrail frameworks, and fail-safe design patterns
Objectives
- Design multi-layer safety systems for AI applications
- Implement guardrail frameworks that prevent harmful outputs
Key concepts
- input validation
- output filtering
- safety classifiers
- hallucination detection
- NeMo Guardrails
- fail-safe patterns
Regulatory Compliance (EU AI Act)
EU AI Act risk categories, documentation requirements, conformity assessments, model cards, data sheets, and building compliance-ready AI systems
Objectives
- Classify AI systems according to EU AI Act risk categories
- Build documentation and audit trails for regulatory compliance
Key concepts
- EU AI Act risk tiers
- conformity assessment
- model cards
- data sheets
- audit trails
- impact assessments
Red-Teaming AI Systems
Systematic adversarial testing of AI systems, attack taxonomies, jailbreak testing, automated red-teaming, and building robust defense layers
Objectives
- Conduct systematic red-teaming exercises for AI systems
- Design defense-in-depth strategies against adversarial attacks
Key concepts
- attack taxonomies
- jailbreak testing
- automated red-teaming
- adversarial robustness
- defense-in-depth
- vulnerability disclosure
7 · Expert ML Engineering5 topics
ML Platform Engineering
Building and operating ML platforms using Kubeflow, Vertex AI, or SageMaker for scalable training and serving
Objectives
- Design and operate end-to-end ML platforms for large engineering teams
- Evaluate trade-offs between managed ML platforms and custom infrastructure
Key concepts
- Kubeflow Pipelines
- Vertex AI Pipelines
- SageMaker
- pipeline orchestration
- experiment tracking
- distributed training infrastructure
- serving infrastructure
Model Governance & Compliance Frameworks
Implementing model governance: model cards, approval workflows, risk tiering, audit trails, and regulatory compliance
Objectives
- Design a model governance framework covering the full ML lifecycle
- Map model risk tiers to regulatory obligations and approval workflows
Key concepts
- model cards
- risk tiering
- approval workflows
- audit trails
- model inventory
- EU AI Act mapping
- responsible AI frameworks
Large-Scale Distributed Training
Scaling ML model training with data parallelism, model parallelism, pipeline parallelism, and gradient checkpointing
Objectives
- Choose the right parallelism strategy for billion-parameter models
- Configure ZeRO optimizer stages and mixed precision for memory efficiency
Key concepts
- data parallelism
- model parallelism
- pipeline parallelism
- ZeRO optimizer
- gradient accumulation
- mixed precision
- DeepSpeed/FSDP
ML Cost Optimization
Optimizing ML infrastructure costs: spot training, quantization, pruning, knowledge distillation, and serving efficiency
Objectives
- Reduce ML training and serving costs using quantization and pruning
- Design cost-aware inference pipelines that meet latency SLAs
Key concepts
- spot instance training
- INT8 quantization
- model pruning
- knowledge distillation
- batch inference optimization
- GPU utilization
- right-sizing
Foundation Model Fine-Tuning
Fine-tuning foundation models using LoRA, QLoRA, instruction tuning, and RLHF for domain-specific applications
Objectives
- Select and apply PEFT methods for domain adaptation with constrained compute
- Design evaluation harnesses to validate fine-tuned model quality
Key concepts
- LoRA
- QLoRA
- PEFT
- instruction tuning
- RLHF
- DPO
- catastrophic forgetting
- eval harness
Levels: Junior, Mid, Senior, and Expert
Explain it in your own words
Graded by AI · no account · 1 questionNo options to guess from: you write, the AI grades against the topic's rubric and tells you what level your answer reads at.
Already studying AI/ML Engineer?
The assessment measures topic by topic, and your plan skips what you already know.
Open to read. An account is only needed to be measured.
Frequently asked questions
Where should I start studying AI/ML Engineer?
Start with Data & Feature Engineering: Data Collection & Preprocessing, Feature Engineering Techniques, and Feature Stores. The map orders topics by what they depend on, so each one prepares the next.
How long does a full AI/ML Engineer assessment take?
A full assessment of this path takes about 65 minutes. Study time depends on where you start; a diagnostic shows which topics you can skip.
Do I need an account to use this page?
No. The map, objectives and key concepts are open. You only need an account to be measured, get your gap by topic and receive a plan.