Plain-English definitions of the AI and engineering terms I use across this site — no jargon for its own sake.
A Large Language Model predicts the most likely next token given the preceding text. Trained on massive corpora, it can write, summarise, translate and reason within limits. It has no memory between calls beyond what you put in the prompt, and it can be confidently wrong — which is why verification matters.
LLMs process text in tokens, not characters or words. A token is often a word-piece; in English, ~4 characters or ~0.75 words. Pricing and context limits are counted in tokens, so prompt length directly affects cost and latency.
Prompt engineering is the craft of structuring instructions, examples and context so a model produces useful, consistent output. It includes role framing, few-shot examples, output-format constraints and decomposition. It is a real skill, not a magic phrase.
RAG retrieves relevant chunks (usually from a vector database) and injects them into the prompt so the model answers from your data instead of only its training. It reduces hallucination and lets you answer questions about private or fresh content without retraining.
An embedding turns text into a vector of numbers where semantic similarity becomes geometric closeness. Store embeddings in a vector database and you can search by meaning, not keywords — the retrieval half of RAG.
The context window is the token budget for everything the model sees in one call — system prompt, history, retrieved docs and the question. Exceed it and older content is truncated. Bigger windows help, but relevant beats large.
Temperature scales how much randomness the model uses when sampling the next token. Near 0 gives repeatable, conservative output (good for extraction and code); higher values increase variety and creativity (good for brainstorming) at the cost of consistency.
A hallucination is fluent, plausible output that is simply wrong — invented facts, APIs or citations. It happens because the model predicts likely text, not truth. Mitigate with retrieval (RAG), tighter prompts, lower temperature, and — always — human verification of anything that matters.
An AI agent wraps a model in a loop where it can call tools (search, code execution, APIs), observe results and decide the next step toward a goal. More capable than a single answer, and more risky — actions need guardrails, limits and review.
MCP is a protocol that lets AI applications expose tools, resources and prompts to models in a consistent way — so a client can talk to many servers (files, databases, APIs) without bespoke glue for each. Think of it as a common plug for model context.
An API gateway sits in front of one or more services and handles cross-cutting concerns: authentication, routing, rate limiting, metering and response shaping. My Vensix project is an LLM gateway — one key that routes to several model providers and meters cost.
An idempotent request produces the same end state no matter how many times it runs. It is essential for payments and webhooks, where retries are normal — an idempotency key lets the server apply a duplicate as a harmless no-op instead of charging twice.
Rate limiting caps requests per client over time (e.g. 60/minute) to prevent abuse and overload. A rate-limited response returns HTTP 429. Common algorithms include fixed window, sliding window and token bucket.
A webhook is an HTTP callback — instead of you polling, a provider POSTs to your endpoint when something occurs (a payment settles, a build finishes). Treat webhooks as untrusted and retried: verify signatures, and make handlers idempotent.
A JWT encodes claims (who you are, what you can do, when it expires) and is signed so the server can trust it without a session store. Great for stateless auth; just remember tokens can’t be un-issued easily, so keep lifetimes short and validate signatures.
QRIS (Quick Response Code Indonesian Standard) unifies QR payments so a single code works across banks and e-wallets. For Indonesian products it is the default way to take money — no card required. The engineering discipline is server-side verification and idempotency.
Edge computing executes your code in data centres near the visitor, cutting latency and removing the need to pick a region. Platforms like Cloudflare Workers run request-scoped functions at the edge — no server to manage, cost that scales to near-zero at low traffic.
Serverless means you ship code, not machines. The platform handles provisioning, scaling and idle. You pay per use and design around short-lived, stateless execution. Trade-offs: cold starts, execution limits, and long jobs belong in queues, not the request.
A CDN caches and serves static assets (and increasingly dynamic responses) from edge locations worldwide, so users get bytes from nearby instead of a distant origin. It cuts latency, absorbs traffic spikes and reduces origin load.
CI runs your build and tests automatically on every change; CD ships the result to staging or production with minimal manual steps. Done well, it gates releases behind tests (my CoalTrack line runs 400+ tests before a signed build) so shipping is boring and safe.
Row-Level Security attaches policies to a table so it only returns rows a user is allowed to see, no matter which query asks. It moves authorization from scattered app code into the data itself — a forgotten filter can’t leak anything, because the database refuses.
The OWASP Top 10 ranks the most critical web-app security risks (injection, broken access control, and so on). It is the baseline checklist for building and reviewing web apps — know it, and design against each item.
SQL injection exploits input that gets concatenated into a query, letting an attacker run their own SQL. The fix is parameterised queries (never string-building SQL with user input) plus least-privilege database accounts.
An append-only audit trail guarantees history can’t be quietly rewritten. In CoalTrack, PostgreSQL triggers reject UPDATE and DELETE on attendance and approval tables, so the app can add facts but not alter them — essential when payroll or a dispute depends on the record.
No terms match that search.