From Model Math to Production AI
Follow the quantities that shape an AI system: parameters create compute and memory demand, context and caching govern serving, embeddings power retrieval, and usage becomes cost, latency and reliability.
Models & Compute
Connect architecture size to operations, memory, training time and optimization choices.
Model Parameters
Count trainable and non-trainable model weights.
Neural Network FLOPs
Estimate arithmetic work for a network.
LLM Training Compute
Estimate training operations from parameters and tokens.
Parameter Memory
Translate weights into storage and VRAM needs.
Quantization
Estimate model size after precision reduction.
Distributed Training Time
Estimate runtime across multiple accelerators.
LLM Memory & Serving
See how context, cache, batch size and throughput determine runtime capacity.
Context Window
Check whether a request fits within model context.
Batch Size vs Sequence
Measure tokens processed together.
KV Cache Memory
Estimate attention-cache memory during generation.
LLM GPU VRAM
Estimate total accelerator memory required.
Tokens per Second
Measure generation throughput.
Serving Concurrency
Estimate simultaneous requests a deployment can serve.
RAG & Embeddings
Turn documents into vectors, retrieve relevant chunks and measure retrieval quality and context cost.
Embedding Similarity
Measure semantic alignment between vectors.
Embedding Storage
Estimate vector-index data size.
RAG Chunk Size
Plan document segmentation for retrieval.
Retrieval Metrics
Measure retrieval precision, recall and ranking quality.
RAG Context Cost
Estimate the prompt cost added by retrieved content.
MCP Tool Schema Tokens
Estimate context consumed by tool definitions.
AI Cost & Reliability
Convert tokens and workloads into cost, latency, savings and end-to-end system success.
Token Count
Estimate billable input and output volume.
AI API Cost
Calculate usage-based model charges.
Inference Latency
Estimate end-to-end response time.
Prompt Caching Savings
Estimate savings from reusing prompt prefixes.
AI Agent Cost
Estimate multi-step agent execution cost.
Agent Reliability
Estimate end-to-end success across dependent steps.