AI concepts I use to build deployable systems.

Artificial Intelligence → Generative AI

AI is the broad field of building systems that perform tasks associated with human intelligence. Machine learning learns patterns from data; deep learning uses layered neural networks; foundation models are trained broadly and adapted to many downstream tasks.

Deployment lens

The useful question is not “is this AI?” but “what decision or workflow does the system improve, under what constraints, and how will we know it works?”

LLMs are predictive systems with application-level behavior

A large language model predicts the next token from context. Transformer architectures make this possible through attention and feed-forward layers, while parameters store learned statistical structure from training.

What matters in an application

• Model capability and instruction-following reliability
• Context capacity and the cost of long prompts
• Tool schema compatibility and output constraints
• Latency, throughput, privacy, and deployment cost

Context is a system budget, not an abstract maximum

Tokens are the units a model reads and generates. The context window includes system instructions, tool definitions, conversation history, retrieved content, and the current task. A model can fit in memory while the agent still fails because its effective context budget is too small.

FDE debugging question

When an agent breaks, inspect the full request and protocol—not only the model size. Context overhead, serialization, and repeated tool results can be the real bottleneck.

Prompting is part of the interface contract

A good system prompt defines role, boundaries, tool usage, output shape, and failure behavior. Prompt quality matters, but reliable applications also need validation, retries, observability, and a clear path when the model cannot complete the task.

Retrieval grounds a model in a changing knowledge source

Embeddings represent text as vectors so semantically related content can be retrieved. A RAG system typically chunks source material, embeds it, retrieves candidates, and supplies selected evidence to the model.

In enterprise settings, retrieval quality, access control, freshness, citation behavior, and fallback handling matter as much as the vector database.

An agent is a controlled loop around a model

An agent observes a task, decides whether to call a tool, receives the result, and continues until it reaches a stopping condition. Tool calling turns language output into actions, so the tool boundary must be explicit and least-privileged.

Safety baseline

Use allowlists, path restrictions, confirmation gates, timeouts, structured arguments, and audit logs for tools that touch files, shells, data, or external systems.

Local deployment exposes the real constraints

Ollama and similar runtimes make local models accessible, but usable deployment depends on memory, quantization, context length, prompt processing speed, generation speed, thermal limits, and protocol compatibility.

Evaluation is evidence for a deployment decision

Start with representative tasks and define success before tuning. Test functional correctness, tool behavior, latency, context stress, failure recovery, safety boundaries, and repeatability. A demo is evidence of possibility; an evaluation is evidence of readiness.

Enterprise AI begins with workflow discovery

Deployment work connects domain context, customer pain, technical scope, architecture, integration, and change management. The goal is a system that fits an existing operating model—not a generic model bolted onto a process.