Artificial Intelligence & Data Science · 1.4

AI/ML Engineer

Build, optimize, and deploy production-grade AI/ML systems at scale.

9–12 months6 phases26 skills

Foundations

Python masteryadvanced OOP (abstract base classes, mixins, dataclasses), design patterns (factory, strategy, observer), async programming (asyncio, aiohttp for concurrent API calls), type hints with Pydantic for runtime data validation, code formatting with black and linting with ruff, profiling with cProfile and memory_profiler

Mathematics

Mathematicslinear algebra, calculus, optimization theory (convex optimization, gradient descent variants), probability — same depth as Data Scientist Phase 1

Data structures and algorithms

Data structures and algorithmsarrays, hash maps, trees, graphs, dynamic programming, complexity analysis (Big-O) — enough for coding interviews and efficient pipeline design

Software engineering for ML

Software engineering for MLGit workflows for experiments (feature branches), Cookiecutter Data Science project structure (src/data, src/features, src/models, tests/, configs/), configuration management with Hydra (YAML configs with CLI overrides), testing ML code with pytest (test data processing, test model inference, test training convergence), CI with GitHub Actions

Phase 02 · 11h

Machine Learning Engineering

ML fundamentals

ML fundamentalssupervised/unsupervised/reinforcement learning, bias-variance tradeoff, cross-validation strategies — same depth as Data Scientist Phase 3

Scikit-learn advanced

Scikit-learn advancedpipelines with custom estimators (subclass BaseEstimator and TransformerMixin), column transformers for heterogeneous data, model selection with cross_val_score and GridSearchCV/Optuna

Gradient boosting mastery

Gradient boosting masteryXGBoost (hyperparameter tuning — n_estimators, max_depth, learning_rate, subsample, colsample_bytree, reg_alpha, reg_lambda; early stopping with eval_set; feature importance with plot_importance), LightGBM (leaf-wise growth, handling categoricals with cat_features parameter), CatBoost (native categorical handling, ordered target encoding)

Feature engineering at scale

Feature engineering at scaleFeast feature store (define FeatureViews from data sources, materialize offline→online store, retrieve features for training with get_historical_features and for serving with get_online_features)

Experiment tracking

Experiment trackingWeights & Biases (wandb.init, wandb.log for metrics/charts/confusion matrices, hyperparameter sweeps with wandb.sweep using Bayesian optimization, artifact versioning for datasets and models, model registry, shareable reports) for deep learning; MLflow for simpler projects and self-hosted requirements

Phase 03 · 42h

Deep Learning & Frameworks

PyTorch advanced

12h

PyTorch advancedcustom Datasets and DataLoaders (num_workers, pin_memory, custom collate_fn for variable-length sequences), mixed precision training with torch.cuda.amp (autocast + GradScaler for 2x speed, half memory), gradient accumulation for simulating larger batch sizes, gradient clipping (clip_grad_norm_ to prevent exploding gradients), distributed training with DistributedDataParallel (DDP) for multi-GPU, debugging with detect_anomaly() and learning rate finder

Computer vision

Computer visionmodern architectures (Vision Transformer/ViT, ConvNeXt, EfficientNet), object detection with YOLO v8 (pip install ultralytics — train custom detector with model.train(data='dataset.yaml', epochs=100)), image segmentation with SAM (Segment Anything Model), data augmentation with Albumentations (RandomResizedCrop, HorizontalFlip, ColorJitter, Normalize, ToTensorV2)

NLP and LLMs

14h

NLP and LLMsencoder models (BERT, RoBERTa, DeBERTa) fine-tuned for classification/NER/QA, decoder models for generation (Llama 3, Mistral 7B), fine-tuning LLMs with LoRA (Low-Rank Adaptation — freeze pretrained weights, add small trainable matrices to attention layers with PEFT library: LoraConfig(r=16, lora_alpha=32, target_modules=["q_proj","v_proj"])), QLoRA (4-bit quantization with bitsandbytes + LoRA — enables fine-tuning 7B model on 8GB VRAM), training with Hugging Face TRL library (SFTTrainer for supervised fine-tuning), evaluation with perplexity, BLEU, ROUGE, and LLM-as-judge

Generative AI

Generative AIdiffusion models (Stable Diffusion architecture overview), GANs (StyleGAN concepts), VAEs — understanding the landscape

Model optimization

Model optimizationquantization (post-training with torch.quantization.quantize_dynamic, GPTQ and AWQ for LLMs), knowledge distillation (train smaller "student" to mimic larger "teacher"), pruning (remove near-zero weights), ONNX export (torch.onnx.export for cross-platform optimized inference with ONNX Runtime)

Phase 04 · 10h

LLM Engineering

RAG (Retrieval-Augmented Generation) systems

RAG (Retrieval-Augmented Generation) systemsdocument loading and chunking (RecursiveCharacterTextSplitter from LangChain, semantic chunking), embedding models (sentence-transformers all-MiniLM-L6-v2 for local, OpenAI text-embedding-3-small for API), vector databases (Pinecone for managed, Chroma for lightweight prototyping, pgvector for PostgreSQL-native), retrieval pipeline (embed query → search vector DB → inject top-k chunks into LLM prompt → generate answer with citations), evaluation with RAGAS framework (context relevance, answer faithfulness, answer relevance), advanced RAG (hybrid search combining vectors + BM25, cross-encoder re-ranking, query expansion)

AI agents and tool use

AI agents and tool usefunction calling with tool schemas (JSON describing name, description, parameters), LangChain agents with @tool decorator, LangGraph for stateful multi-step agents (state graphs with nodes, edges, cycles, checkpointing, human-in-the-loop), multi-agent orchestration with CrewAI, safety guardrails (preventing harmful tool use, output validation, cost controls)

LLM frameworks

LLM frameworksLangChain (chains, agents, tools, memory), LlamaIndex (data ingestion, indexing, querying)

Evaluation

EvaluationBLEU, ROUGE, BERTScore, human evaluation design, LLM-as-judge rubric prompting, red-teaming for safety

Phase 05 · 14h

MLOps & Production Systems

Model serving

Model servingFastAPI wrapping a model (simplest approach), TorchServe (model packaging with torch-model-archiver, worker management, dynamic batching), NVIDIA Triton Inference Server (supports PyTorch/ONNX/TensorRT, dynamic batching, model ensemble, GPU scheduling), vLLM for high-throughput LLM serving (PagedAttention for efficient memory management)

Containerization

ContainerizationDocker for ML (GPU support with NVIDIA Container Toolkit, CUDA base images), Kubernetes for ML workloads (GPU scheduling with NVIDIA device plugin, Jobs for training, Deployments for serving, HPA auto-scaling, Helm charts for packaging)

CI/CD for ML

CI/CD for MLon code change → run tests → train on sample data → evaluate → if metrics pass → build Docker image → deploy to staging → integration tests → canary deployment to production → monitor → full rollout

Monitoring

Monitoringdata drift and model drift detection with Evidently AI (scheduled comparison reports), infrastructure monitoring with Prometheus + Grafana (latency p50/p95/p99, GPU utilization, error rates), business metric tracking, automated alerting via PagerDuty or Slack

Cloud ML platforms

Cloud ML platformsAWS SageMaker (Training Jobs, Endpoints with auto-scaling, Model Registry), GCP Vertex AI (equivalent managed services); GPU instance selection (T4 for inference, A10G for balanced, A100/H100 for heavy training), spot instances with checkpointing for 70-90% cost savings

Phase 06 · 8h

Capstone

Capstone Project 1

Capstone Project 1Production RAG system — document ingestion pipeline (S3 → chunk → embed → Pinecone), FastAPI backend with query endpoint, React or Streamlit chat frontend, RAGAS evaluation on 50 Q&A pairs, monitoring dashboard, deployed with Docker Compose

Capstone Project 2

Capstone Project 2End-to-end ML pipeline — feature engineering → Feast feature store → model training with W&B tracking → Optuna tuning → ONNX optimization → Triton serving → Evidently monitoring → Prometheus/Grafana dashboards

Capstone Project 3

Capstone Project 3Fine-tuned LLM — QLoRA fine-tune Llama 3 8B on domain-specific data using TRL SFTTrainer, evaluate with automated + human metrics, merge LoRA weights, quantize with AutoGPTQ, serve with vLLM, write model card documenting training data, performance, and limitations

Ready to start AI/ML Engineer?

Start this roadmap to track progress and generate courses for each skill. Login required only to begin.