ml-data-engineering

name

ml-data-engineering

description

Use for ML model training, inference serving (vLLM, TorchServe, ONNX), data pipelines, RAG/LLM systems, and MLOps. Covers classification, regression, neural networks, ETL, feature engineering, vector databases, model deployment, experiment tracking. PyTorch, scikit-learn, HuggingFace, MLflow, W&B, FAISS, pgvector, Chroma, LoRA, QLoRA, PEFT, DVC, Delta Lake, Spark, Feast, Evidently, data drift, A/B testing.

ML & Data Engineering

Master skill for machine learning, data science, and data engineering. Routes to specialized references by task type.

Task Router

Task	Reference	When to use
Data pipelines, pandas, polars, ETL, validation	references/data-engineering.md	Data loading, cleaning, transformation, schema validation
Classification, regression, clustering	references/classical-ml.md	scikit-learn, feature engineering, model selection
Neural networks, GPU training, CNNs, transformers	references/deep-learning.md	PyTorch, transfer learning, training loops
RAG, embeddings, prompt engineering, LLM eval	references/llm-patterns.md	LLM integration, chain-of-thought, few-shot
Model serving, experiment tracking, CI/CD for ML	references/mlops.md	MLflow, W&B, deployment pipelines
pgvector, Pinecone, Chroma, FAISS	references/vector-databases.md	Embedding storage and retrieval
Model comparison, statistical significance	references/ml-experimentation.md	Hypothesis testing, nested CV, reproducibility
Hybrid retrieval, reranking, HyDE, RAPTOR	references/advanced-rag.md	Multi-vector RAG, agentic RAG
TorchServe, BentoML, vLLM, ONNX Runtime	references/model-serving.md	Production inference, batching, GPU serving
Data drift, model monitoring, Evidently	references/monitoring-drift.md	PSI, concept drift, prediction distribution
LoRA, QLoRA, adapter tuning, HuggingFace PEFT	references/peft-fine-tuning.md	Parameter-efficient fine-tuning
DVC, Delta Lake, dataset lineage	references/data-versioning.md	Reproducible data pipelines
A/B testing, canary rollout, shadow mode	references/online-evaluation.md	Traffic splitting, sequential testing

Decision Trees

Classical ML vs Deep Learning: Tabular <100k rows → classical (gradient boosting). Images/audio/video → deep learning. Text <10k examples → fine-tuned transformer or TF-IDF + classical. Text generation → LLM.

Fine-tune vs Prompt: Try prompting first. Fine-tune when: consistent format needed, domain knowledge, latency matters, cost at scale. RAG before fine-tuning for knowledge tasks.

Framework: Tabular → scikit-learn + XGBoost/LightGBM. Deep learning → PyTorch. LLM fine-tuning → HuggingFace + PEFT. LLM inference → vLLM or API.

Core Principles

1. Data quality > model complexity — clean data with logistic regression beats dirty data with transformers
2. Reproducibility — pin seeds, pin versions, track experiments (MLflow/W&B)
3. Start simple — baseline with simple model, justify complexity with metrics
4. Validate like production — time-based splits for time series, stratified for imbalanced
5. Version your data — DVC, Delta Lake, or checksums
6. Fail fast on data issues — validate schemas at ingestion (pandera, Great Expectations)

Anti-Patterns

Anti-pattern	Fix
Training on test data	Strict train/val/test split before preprocessing
Data leakage	Fit scalers on train only
No baseline	Start with simple model
Ignoring class imbalance	Use F1/AUROC, stratified splits
Notebook spaghetti	Extract to .py modules
Deploying without monitoring	Track prediction distributions, drift

Layer 3 Skills

Skill	Use when
data-engineering	ETL pipelines, data validation, pandas/polars
classical-ml	scikit-learn, feature engineering, model selection
deep-learning	PyTorch, neural networks, GPU training
llm-patterns	RAG, embeddings, prompt engineering
mlops	Experiment tracking, model deployment, CI/CD
vector-databases	Embedding storage, similarity search
advanced-rag	Hybrid retrieval, reranking, agentic RAG
model-serving	Production inference, TorchServe, vLLM
monitoring-drift	Data drift detection, model monitoring
peft-fine-tuning	LoRA, QLoRA, adapter tuning
data-versioning	DVC, Delta Lake, dataset lineage
online-evaluation	A/B testing, canary rollout, shadow mode