--- title: "Glossary: ML infrastructure and agent terms" description: "Plain definitions of the 28 terms that the case studies, posts, and developer docs on sudhanva.me use." canonical: https://sudhanva.me/glossary/ last_updated: 2026-11-11 --- # Glossary: ML infrastructure and agent terms The case studies, posts, and developer docs on this site use these terms. ## A2A (Agent2Agent) An open protocol that lets one AI agent find and call another agent over HTTP. An agent publishes an Agent Card that lists its skills and endpoints. See [A2A Agent Card](https://sudhanva.me/.well-known/agent-card.json). ## ARD (Agentic Resource Discovery) A JSON manifest at /.well-known/ard.json that lists the MCP servers and agents of a site, so an agent can find them in one request. See [ARD manifest](https://sudhanva.me/.well-known/ard.json). ## Argo CD A GitOps tool that reads Kubernetes manifests from a Git repository and applies them to a cluster until the cluster matches Git. See [K3s on Oracle Cloud](https://sudhanva.me/work/k3s-oracle/). ## Batch inference A job that runs a trained model over a large, fixed dataset at once, not on single live requests. Throughput and cost per row matter more than latency. See [Billion-row batch inference](https://sudhanva.me/work/case-studies/billion-row-batch-inference/). ## Content Signals robots.txt directives that state whether crawlers may use the content for search, for AI answers, and for model training. See [robots.txt](https://sudhanva.me/robots.txt). ## DDP (DistributedDataParallel) A PyTorch module that runs one copy of a model on each GPU. After each backward pass, it averages the gradients across all GPUs, so every copy stays the same. See [PyTorch DDP for fine-tuning Gemma 4](https://sudhanva.me/blog/posts/pytorch-ddp-fine-tuning-gemma-4/). ## DNS-AID DNS for AI Discovery: SVCB or HTTPS records under the _agents label of a domain that tell agents where the agent endpoints of the domain are. ## Flyte A workflow orchestration platform for data and ML pipelines. It runs each task as a container on Kubernetes and records the inputs and outputs of each run. See [Kubernetes ML platform](https://sudhanva.me/work/case-studies/production-ml-platform-kubernetes/). ## GitOps A way to operate infrastructure in which a Git repository holds the wanted state, and a controller changes the live system to match it. ## GPU autoscaling A cluster adds GPU nodes when jobs wait for GPUs and removes idle nodes when the jobs finish. The cluster then pays only for the GPUs in use. See [Kubernetes ML platform](https://sudhanva.me/work/case-studies/production-ml-platform-kubernetes/). ## Idempotency key A unique value that a client sends with a create request. If the client sends the same request again, the server returns the first result and does not create a second job. See [Profile insights](https://sudhanva.me/docs/profile-insights/). ## IndexNow A protocol that lets a site tell search engines such as Bing which URLs changed, so they crawl those URLs again. ## K3s A small Kubernetes distribution packaged as one binary. It runs well on single machines, edge devices, and small clouds. See [K3s on Oracle Cloud](https://sudhanva.me/work/k3s-oracle/). ## Kubernetes A container orchestrator. It schedules containers onto machines, restarts them when they fail, and scales them up and down. See [Bare-metal homelab](https://sudhanva.me/work/homelab/). ## llms.txt A Markdown file at the root of a site that gives language models a short index of the site. llms-full.txt holds the full text. See [llms.txt](https://sudhanva.me/llms.txt). ## LoRA (Low-Rank Adaptation) A fine-tuning method that freezes the base model weights and trains small low-rank adapter matrices. It trains a small fraction of the parameters and needs much less GPU memory. See [PyTorch DDP fine-tuning study](https://sudhanva.me/work/case-studies/pytorch-ddp-gemma-4-fine-tuning/). ## MCP (Model Context Protocol) An open protocol that lets AI applications call tools and read resources on a server. This site runs a read-only MCP server and a separate MCP server for profile-insight jobs. See [MCP server guide](https://sudhanva.me/developers/mcp/). ## MCP App An MCP extension that lets a tool return an interactive HTML view, which a compatible host shows inside the conversation. See [MCP server guide](https://sudhanva.me/developers/mcp/). ## MLflow A tool that records ML experiments: the parameters, metrics, and model files of each training run. ## Model serving A trained model runs behind an API and answers live requests. Latency, throughput, and cost per request are the main constraints. See [Real-time transformer prediction](https://sudhanva.me/blog/posts/real-time-transformer-prediction-at-autodesk/). ## NLWeb A protocol that lets a client ask a site a question in natural language at an /ask endpoint and get structured results. See [NLWeb endpoint](https://sudhanva.me/developers/nlweb/). ## OpenAPI A machine-readable description of an HTTP API: its endpoints, parameters, request bodies, and responses. See [OpenAPI document](https://sudhanva.me/openapi.json). ## Ray A framework for distributed Python. It spreads tasks across a cluster and is often used for parallel training, tuning, and batch inference. See [Billion-row batch inference](https://sudhanva.me/work/case-studies/billion-row-batch-inference/). ## RFC 9728 (protected-resource metadata) An OAuth standard. A protected resource publishes a JSON document that names its authorization servers, scopes, and token methods. See [Authentication](https://sudhanva.me/docs/authentication/). ## Terraform An infrastructure-as-code tool. You declare cloud resources in files, and Terraform creates or changes them to match. ## torchrun The PyTorch launcher that starts one training process per GPU and sets the environment variables that DDP needs. See [PyTorch DDP for fine-tuning Gemma 4](https://sudhanva.me/blog/posts/pytorch-ddp-fine-tuning-gemma-4/). ## Web Bot Auth A way for bots to sign their HTTP requests with a published key, so a site can verify which bot sent a request. ## WebMCP A browser API that lets a web page register tools for an AI agent in the browser, so the agent calls the tools and does not read the page layout. See [WebMCP script](https://sudhanva.me/webmcp.js). ## Sitemap - [Site map in Markdown](https://sudhanva.me/sitemap.md): every page on sudhanva.me, grouped by section