ML Building Blocks & Models
A comprehensive reference of every component you need to design production ML systems — from data ingestion and feature engineering to deployment, evaluation, RAG and LLM ops.
Batch Data Source
Historical data from databases, data lakes, or files
Streaming Source
Real-time data from Kafka, Kinesis, or event streams
API Endpoint
REST or GraphQL API for data ingestion
File Upload
Upload files (CSV, JSON, Parquet, images)
Webhook
Receive data via HTTP webhooks
Web Scraper
Extract structured data from websites for ML training and RAG pipelines
Data Validation
Schema validation, null checks, outlier detection
Data Cleaning
Handle missing values, remove duplicates, normalize
Data Transformation
ETL transformations, aggregations, joins
Normalization
Min-max, Z-score, or decimal scaling
Deduplication
Remove duplicate records
Feature Extraction
Extract features from raw data (embeddings, encodings)
Feature Selection
Select relevant features, reduce dimensionality
Feature Store
Centralized repository for feature management
Encoding
One-hot, label, or target encoding
Feature Scaling
StandardScaler, MinMaxScaler, RobustScaler
Train/Test Split
Split data into training, validation, and test sets
Model Training
Train ML model with hyperparameter tuning
Hyperparameter Tuning
Grid search, random search, or Bayesian optimization
Cross-Validation
K-fold, stratified, or time-series cross-validation
Full Fine-tuning
Update all model parameters on task-specific data
LoRA
Low-Rank Adaptation - add small trainable matrices to attention layers
QLoRA
Quantized LoRA - 4-bit quantization + LoRA for memory efficiency
Adapter Layers
Insert small trainable modules between frozen transformer layers
Prefix Tuning
Learn continuous soft prompts prepended to each layer
Prompt Tuning
Learn task-specific prompt embeddings (input layer only)
IA³
Infused Adapter - learn rescaling vectors for activations
Instruction Tuning
Fine-tune on instruction-following datasets (e.g., Alpaca, ShareGPT)
RLHF
Reinforcement Learning from Human Feedback with reward model
Reward Modeling
Train reward model from human preference comparisons
DPO
Direct Preference Optimization - simplified RLHF without reward model
ORPO
Odds Ratio Preference Optimization - single-stage SFT + preference
Constitutional AI
Self-improvement via AI feedback based on constitutional principles
Feature Extraction
Freeze base model, train only the classification head
Domain Adaptation
Adapt pretrained model to a new domain (e.g., medical, legal)
Continued Pretraining
Further pretrain on domain-specific corpus before fine-tuning
Knowledge Distillation
Train smaller student model to mimic larger teacher model
Multi-Task Learning
Train on multiple tasks simultaneously with shared representations
Transfer Learning
Reuse pretrained model knowledge for new tasks via fine-tuning or feature extraction
Active Learning
Iteratively select most informative samples for labeling to minimize annotation cost
Model Quantization
Reduce model precision (FP32→INT8/INT4) for faster inference and smaller footprint
Accuracy
Overall classification accuracy (TP+TN)/(Total)
Precision/Recall/F1
Precision, Recall, F1-Score (per class & macro/micro)
Confusion Matrix
Visualize TP, TN, FP, FN across classes
ROC-AUC Curve
Receiver Operating Characteristic & Area Under Curve
PR Curve
Precision-Recall curve (for imbalanced datasets)
Log Loss
Cross-entropy loss for probabilistic predictions
Cohen's Kappa
Agreement metric accounting for chance
MAE
Mean Absolute Error
MSE / RMSE
Mean Squared Error / Root MSE
R² Score
Coefficient of determination
MAPE
Mean Absolute Percentage Error
Residual Plot
Visualize prediction residuals
Precision@K
Precision at top K results
Recall@K
Recall at top K results
MAP
Mean Average Precision
MRR
Mean Reciprocal Rank
NDCG
Normalized Discounted Cumulative Gain
Hit Rate
Fraction of queries with at least one relevant result
Catalog Coverage
Percentage of items ever recommended
Diversity Score
Intra-list diversity of recommendations
Novelty Score
Average popularity rank of recommended items
Serendipity
Unexpected but relevant recommendations
CTR / Conversion
Click-through rate & conversion metrics
BLEU Score
Bilingual Evaluation Understudy (translation/generation)
ROUGE Score
Recall-Oriented Understudy (summarization)
BERTScore
Semantic similarity using BERT embeddings
Perplexity
Language model quality metric
Faithfulness
Factual consistency with source (RAG)
Answer Relevance
Relevance of generated answer to query (RAG)
IoU / Jaccard
Intersection over Union (detection/segmentation)
mAP (Detection)
Mean Average Precision for object detection
Dice Coefficient
Segmentation overlap metric
PSNR / SSIM
Image quality metrics (generation/super-resolution)
FID Score
Fréchet Inception Distance (generative models)
Silhouette Score
Cluster cohesion and separation
Davies-Bouldin Index
Cluster similarity ratio
Calinski-Harabasz
Variance ratio criterion
ARI / NMI
Adjusted Rand Index / Normalized Mutual Info
A/B Test Runner
Statistical A/B test framework
Statistical Significance
P-value and confidence interval calculator
Uplift Model
Incremental impact measurement
K-Means Clustering
Partition-based clustering algorithm minimizing within-cluster variance
PCA (Principal Component Analysis)
Dimensionality reduction via eigendecomposition of covariance matrix
Elbow Method
Heuristic for selecting optimal K in clustering via inertia curve
Context Recall
RAG evaluation metric measuring retrieval completeness against ground truth
Linear Regression
Fundamental regression model fitting linear relationships with OLS
Gradient Boosting (XGBoost/LightGBM)
Ensemble method building sequential trees on residual errors
Gaussian Generator
Generate data from Gaussian/Normal distribution
GAN Data Generator
Generative Adversarial Network for synthetic data
VAE Generator
Variational Autoencoder for data generation
Diffusion Generator
Diffusion model for high-quality synthetic data
CTGAN
Conditional Tabular GAN for structured data
TVAE
Tabular VAE for synthetic tabular data
Copula Generator
Copula-based synthetic data preserving correlations
Faker Generator
Rule-based fake data (names, addresses, etc.)
Time Series Generator
Synthetic time series (ARIMA, seasonal patterns)
LLM Data Generator
Use LLMs to generate synthetic training data
SMOTE
Synthetic Minority Over-sampling Technique
SMOTE-NC
SMOTE for mixed numerical/categorical data
Borderline-SMOTE
SMOTE focusing on borderline samples
ADASYN
Adaptive Synthetic Sampling
Random Oversampler
Simple random duplication of minority class
Random Undersampler
Random removal from majority class
Tomek Links
Remove Tomek links from majority class
Edited Nearest Neighbors
Remove samples misclassified by k-NN
Cluster Centroids
Replace majority class with cluster centroids
NearMiss
Heuristic undersampling based on distance
SMOTE + ENN
SMOTE followed by Edited Nearest Neighbors cleaning
SMOTE + Tomek
SMOTE followed by Tomek links removal
Image Augmentation
Rotate, flip, crop, color jitter, mixup, cutout
Text Augmentation
Synonym replacement, back-translation, EDA
Audio Augmentation
Time stretch, pitch shift, noise injection
Mixup
Convex combination of training examples
CutMix
Cut and paste patches between images
Differential Privacy
Add noise for differential privacy guarantees
Federated Synthesis
Generate synthetic data in federated setting
Condensed Nearest Neighbour (CNN)
Undersampling by finding minimal consistent subset preserving 1-NN boundary
Batch Normalization
Normalize layer inputs across mini-batch for stable deep learning training
Model Registry
Version and manage trained models
Model Serving
Deploy model for real-time or batch inference
Load Balancer
Distribute inference requests across replicas
Canary Deploy
Gradual rollout with traffic splitting
Blue-Green Deploy
Zero-downtime deployment with instant switchover
Batch Inference
Run ML predictions on large datasets offline in batch mode
Pipeline Scheduler
Schedule and orchestrate ML pipelines (Airflow, Kubeflow)
CI/CD Pipeline
Automated testing and deployment pipeline
Workflow Engine
DAG-based workflow orchestration
Event Trigger
Trigger pipelines on events
Experiment Tracker
Log ML experiments — hyperparameters, metrics, artifacts — for reproducibility
Document Loader
Load documents from various sources (PDF, DOCX, web)
Text Chunker
Split documents into chunks (recursive, semantic)
Embedding Model
Generate embeddings from text
Vector Store
Store and retrieve vector embeddings
Semantic Search
Vector similarity search
Hybrid Search
Combine keyword and semantic search
Re-Ranker
Re-rank retrieved results (Cohere, BGE-reranker)
Context Assembler
Assemble context for LLM prompt
BM25 (Lexical Search)
Best Match 25 algorithm for term-frequency based document retrieval
Sparse Retrieval
Term-based retrieval using inverted indexes (TF-IDF, BM25)
Learned Sparse Retrieval (SPLADE)
Neural sparse retrieval with learned term expansion (SPLADE, DeepImpact)
ColBERT (Late Interaction)
Late interaction retrieval with per-token matching via MaxSim
Learning to Rank (LTR)
ML-based ranking combining multiple relevance signals (LambdaMART, RankNet)
Self-RAG
Self-reflective RAG with adaptive retrieval and critique tokens
LLM Generator
Language model that generates answers from retrieved context in RAG
Prompt Template
Define reusable prompt templates
Guardrails
Input/output validation and safety filters
Output Parser
Parse LLM output to structured format
Token Counter
Count tokens for context management
Rate Limiter
Control API request rate
Response Cache
Cache LLM responses for cost savings
Agent Orchestrator
Central controller for agent workflows (ReAct, Plan-Execute)
Tool Executor
Execute agent tools and functions
Memory Store
Short-term and long-term agent memory
Planning Module
Generate and refine action plans
Human-in-Loop
Request human approval or input
ReAct Loop
Reasoning + Acting loop pattern
LangGraph Node
Stateful node in LangGraph workflow
CrewAI Agent
Role-based agent in CrewAI framework
Agent Router
Route tasks to appropriate agents
Shared Memory
Shared context between agents
Task Decomposer
Break complex tasks into subtasks
Agent Supervisor
Coordinate and supervise multiple agents
Image Preprocessor
Resize, normalize, and prepare images
Object Detector
Detect and localize objects in images
Image Classifier
Classify images into categories
Segmentation
Semantic or instance segmentation
OCR
Extract text from images
Face Detection
Detect faces in images
YOLO v8 Nano
ModelUltra-fast object detection
YOLO v8 Medium
ModelBalanced object detection
YOLO v8 XLarge
ModelMost accurate YOLO v8
YOLO v9
ModelState-of-the-art object detection
SAM (Segment Anything)
ModelMeta universal segmentation model
SAM 2
ModelMeta video + image segmentation
CLIP
ModelOpenAI vision-language model
DINOv2
ModelMeta self-supervised vision model
ResNet-50
ModelClassic image classification backbone
EfficientNet-B0
ModelEfficient image classification
ViT-Base
ModelVision Transformer base model
Bias Detector
Detect bias in model predictions
Fairness Checker
Check fairness metrics across groups
Explainer (SHAP)
SHAP values for model explanations
Explainer (LIME)
LIME for local explanations
Privacy Filter
Remove or mask PII from data
Content Moderator
Filter harmful or inappropriate content
GPT-4o
ModelOpenAI flagship multimodal model
GPT-4o Mini
ModelCost-effective OpenAI model for simple tasks
Claude 3.5 Sonnet
ModelAnthropic balanced model for most tasks
Claude 3 Opus
ModelAnthropic most capable model
Gemini Pro
ModelGoogle DeepMind multimodal model
Gemini 2.0 Flash
ModelGoogle fast and efficient model
Llama 3.1 8B
ModelMeta small but capable LLM
Llama 3.1 70B
ModelMeta open-source LLM, excellent reasoning
Llama 3.1 405B
ModelMeta largest open-source LLM
Qwen 2.5 7B
ModelAlibaba small multilingual LLM
Qwen 2.5 72B
ModelAlibaba flagship multilingual LLM
Mistral 7B
ModelMistral AI efficient small model
Mixtral 8x7B
ModelMistral AI Mixture of Experts model
DeepSeek V3
ModelDeepSeek flagship MoE model
Phi-3 Mini
ModelMicrosoft small language model
Phi-3 Medium
ModelMicrosoft medium language model
all-MiniLM-L6-v2
ModelFast and lightweight sentence embedding
BGE-small-en-v1.5
ModelBAAI small embedding model
BGE-base-en-v1.5
ModelBAAI base embedding model
BGE-large-en-v1.5
ModelBAAI large embedding model, excellent for RAG
E5-small-v2
ModelMicrosoft small embedding model
E5-large-v2
ModelMicrosoft large embedding model
GTE-large
ModelAlibaba general text embedding
Instructor-XL
ModelInstruction-following embedding model
text-embedding-3-small
ModelOpenAI small embedding model
text-embedding-3-large
ModelOpenAI large embedding model
Cohere Embed v3
ModelCohere multilingual embedding
Whisper Large v3
ModelOpenAI speech-to-text model, state-of-the-art ASR
TTS-1
ModelOpenAI text-to-speech model
NeRF
ModelNeural Radiance Fields for 3D reconstruction
Instant-NGP
ModelFast neural graphics primitives
3D Gaussian Splatting
ModelFast 3D scene reconstruction
Point-E
ModelOpenAI text-to-3D point cloud
Shap-E
ModelOpenAI text-to-3D mesh generation
GET3D
ModelNVIDIA generative 3D model