ORVINUSAI & Quant, in Plain English
The vocabulary of applied AI and quantitative engineering — defined the way we'd explain it on a discovery call, not the way a textbook would.
The ORVINUS glossary defines 36 terms across four areas: large language models and RAG systems, artificial intelligence and machine learning, quantitative and algorithmic trading, and AI infrastructure and MLOps.
LLMs, RAG & Agents
The vocabulary of modern language-model systems — from tokens to multi-agent orchestration.
Large Language Model (LLM)
A large language model is a neural network trained on massive text corpora to understand and generate human language. LLMs such as GPT, Claude, and Gemini power chatbots, copilots, document analysis, and reasoning systems — and are integrated into products through APIs rather than trained from scratch.
Retrieval-Augmented Generation (RAG)
RAG is an architecture that grounds a language model's answers in your own data: relevant documents are retrieved from a knowledge base and supplied to the model as context before it responds. It is the standard cure for hallucination in enterprise chatbots and document Q&A systems.
Embeddings
Embeddings are numeric vector representations of text (or images) that capture meaning, so that similar content sits close together in vector space. They are the foundation of semantic search, RAG retrieval, recommendation systems, and clustering.
Vector Database
A vector database stores embeddings and answers similarity queries — 'find the content most like this' — at scale. Tools like Pinecone or pgvector make it the retrieval backbone of RAG systems and semantic search features.
Fine-Tuning
Fine-tuning continues a pre-trained model's training on domain-specific examples so it performs better at a specialized task. It is used when prompt engineering hits its ceiling — and is only worth it when measured against a proper evaluation baseline.
Prompt Engineering
Prompt engineering is the craft of designing model instructions — structure, examples, constraints, and context — to get reliable, correct outputs from an LLM. In production systems, prompts are versioned and evaluated like code, not written like chat messages.
Context Window
The context window is the amount of text (measured in tokens) a language model can consider at once, including the conversation, retrieved documents, and instructions. Context limits drive architectural decisions like chunking, summarization, and retrieval strategy.
Token
A token is the unit of text an LLM reads and writes — roughly three-quarters of an English word. API pricing, context limits, and latency are all measured in tokens, which makes token budgeting a real engineering concern in AI products.
Hallucination
A hallucination is a confident but false statement produced by a language model. Production systems mitigate it with retrieval grounding (RAG), constrained output formats, verification steps, and evaluation pipelines that measure factuality rather than assuming it.
AI Agent
An AI agent is a language-model-driven program that can plan, use tools (APIs, databases, browsers), and take actions toward a goal rather than just answering a single prompt. Agents power workflow automation, research assistants, and operational copilots.
Multi-Agent System
A multi-agent system coordinates several specialized AI agents — for example a researcher, a writer, and a reviewer — under an orchestration layer that routes work between them. It outperforms single agents on complex, multi-step processes with branching decisions.
AI Copilot
An AI copilot is an assistant embedded inside a product or workflow that drafts, suggests, and automates alongside a human user — think of a CRM that writes follow-ups or an analytics tool that explains its own charts. Copilots are a leading pattern for adding AI value to existing SaaS.
AI & Machine Learning
Core machine learning concepts behind every model ORVINUS trains and deploys.
Neural Network
A neural network is a machine learning model made of layered, interconnected nodes that learn patterns from data by adjusting connection weights during training. Neural networks underpin everything from price prediction models to LLMs.
Deep Learning
Deep learning is machine learning with many-layered neural networks, capable of learning complex representations directly from raw data — images, text, audio, or market series. It requires more data and compute than classical ML but unlocks tasks those methods cannot solve.
Supervised Learning
Supervised learning trains a model on labeled examples — inputs paired with correct outputs — so it can predict labels for new data. Most business ML (churn prediction, lead scoring, fraud detection) is supervised learning on historical records.
Model Training
Training is the process of fitting a model's parameters to data by iteratively minimizing prediction error. Production-grade training includes dataset versioning, validation splits, and experiment tracking so results are reproducible rather than lucky.
Inference
Inference is running a trained model to make predictions on new data — the 'live' phase of machine learning. Inference latency, throughput, and cost dominate production AI economics, which is why serving infrastructure is engineered separately from training.
Natural Language Processing (NLP)
NLP is the field of AI concerned with understanding and generating human language — classification, extraction, summarization, translation, and sentiment. Modern NLP is largely built on transformer models and powers document intelligence and text analytics.
Sentiment Analysis
Sentiment analysis uses NLP to quantify the emotional tone of text — positive, negative, or neutral — across reviews, news, social media, or support tickets. In finance it feeds trading signals; in products it feeds customer intelligence.
Computer Vision
Computer vision enables software to interpret images and video: recognizing objects, reading documents, measuring, and inspecting. Applications range from product catalog intelligence to OCR-driven data extraction pipelines.
Predictive Analytics
Predictive analytics uses statistical and machine learning models on historical data to forecast future outcomes — demand, churn, prices, risk. The output is a decision input: a number with confidence attached, not a guess.
Quant & Algorithmic Trading
The terms that matter when software trades — strategy, risk, and measurement.
Algorithmic Trading
Algorithmic trading executes trading strategies automatically through software — from signal generation to order placement — using mathematical models instead of manual decisions. It brings speed, discipline, and 24/7 operation that discretionary trading cannot match.
Quantitative Analysis
Quantitative analysis applies mathematics and statistics to financial data to find patterns, price instruments, and manage risk. It is the research discipline behind every serious trading system and market intelligence platform.
Backtesting
Backtesting simulates a trading strategy on historical market data to estimate how it would have performed before risking real capital. Honest backtests model transaction costs, slippage, and look-ahead bias — without those, results are fiction.
Slippage
Slippage is the difference between the price a trade was expected to execute at and the price it actually got. It grows with order size and volatility, and any backtest or execution system that ignores it will overstate real-world returns.
Drawdown
Drawdown is the peak-to-trough decline of a portfolio or strategy — the pain metric. Maximum drawdown tells you the worst losing streak a strategy has endured, and risk management exists chiefly to control it.
Alpha
Alpha is the excess return of a strategy beyond what market exposure alone would deliver — the value the strategy itself adds. Finding, verifying, and protecting alpha is the entire business of quantitative trading.
OHLC
OHLC stands for Open, High, Low, Close — the four prices that summarize a trading period, rendered visually as a candlestick. OHLC series are the standard input to charting, pattern recognition, and most technical strategies.
Options Greeks
The Greeks (delta, gamma, theta, vega, rho) measure how an option's price responds to changes in the underlying price, time, and volatility. They are the control levers of options risk management and derivatives pricing models.
Volatility Modeling
Volatility modeling estimates how much an asset's price fluctuates — historically and implied by options markets — using statistical models. It feeds position sizing, options pricing, and regime detection in trading systems.
Portfolio Optimization
Portfolio optimization allocates capital across assets or strategies to maximize expected return for a chosen level of risk. Modern implementations blend classical mean-variance methods with constraints that reflect real trading: costs, liquidity, and drawdown limits.
AI Infrastructure & MLOps
What it takes to run AI in production: serving, scaling, and observability.
MLOps
MLOps is DevOps for machine learning: versioning models and data, automating training and deployment pipelines, and monitoring model behavior in production. It is the difference between a model that worked once and a model that keeps working.
Model Serving
Model serving is the infrastructure that exposes trained models as reliable, low-latency APIs — handling batching, scaling, and versioned rollouts. Serving is engineered around latency and cost targets, not just correctness.
GPU Acceleration
GPU acceleration runs the parallel math of neural networks on graphics processors, making training and inference orders of magnitude faster than CPUs. Right-sizing GPU infrastructure — cloud vs dedicated, batch vs real-time — is a core cost lever in AI systems.
Auto-Scaling
Auto-scaling automatically adds or removes compute capacity as load changes, so systems stay responsive during spikes and cheap during quiet hours. Cloud-native AI deployments treat it as a default, not an optimization.
Observability
Observability is the ability to understand a system's internal state from its outputs — structured logs, metrics, and traces. For AI systems it extends to model behavior: drift, accuracy decay, and cost per request. You cannot promise 99.9% uptime without it.
CI/CD
Continuous Integration and Continuous Deployment automate the path from code change to production: tests run on every commit, and releases ship through a repeatable pipeline instead of manual steps. It is how engineering teams make deployments boring — and frequent.
Fluent in Both Languages
We translate between business goals and AI engineering every day. Bring the problem in your words — we'll bring the architecture.