Wiki

AI

AI across machine learning, generative AI, agents, production systems, evaluation, infrastructure, and governance.

Artificial intelligence covers systems built from models and data, which also depend on software and workflow design. These systems support decisions, generate content, retrieve knowledge, or automate product work.

In older conversations that often means machine learning, deep learning, and NLP. In newer conversations it often means LLMs, generative AI, retrieval-augmented generation, and AI agents.

The useful boundary isn’t whether a system calls a model, but whether it changes a workflow. Teams also need to evaluate, operate, secure, and improve it. Modern generative AI work connects back to statistical rigor. AI delivery includes human-centered design, and agents sit inside normal software engineering discipline.^[1]

Janna Lipenkova’s Creating Intelligent Products develops the same product-focused view. It treats AI as an end-to-end product discipline, not just model development. The product scope includes market research, UX, and business model fit. Artificial Intelligence with Python by Prateek Joshi covers the broader toolkit that underlies these product systems. The toolkit runs from search and optimization through ML, deep learning, and reinforcement learning.

AI as Product Engineering

AI is product engineering, not standalone prediction, so full-stack engineering and RAG sit in the same skill stack. Knowledge management, LLMOps, and product delivery become delivery concerns.^[2]

That makes AI engineering closer to shipping a data product than to only training a model.

AI engineers combine product discovery and backend delivery. They also work with LLM tooling and decide what should be automated. Projects and skill proof matter more than credentials.^[3]

That’s why the AI page connects to both the AI Engineer Role and the AI Engineering Roadmap.

AI product thinking also shows up before the LLM era. AI product teams work with roadmaps, customer research, business metrics, and MLOps priorities. They work backward from business problems and add success metrics, SLAs, and data quality.^[4]

That discussion links AI to data product management, MLOps, and data engineering.

Machine Learning, LLMs, and Generative AI

Machine learning is the technical base under much of the AI archive. It covers feature work, labels, and training data. It also covers model validation, deployment, monitoring, and tradeoffs between accuracy and interpretability. Use the narrower Machine Learning Engineer Role when the discussion is about supervised learning systems, feature pipelines, or classical production ML.

The newer LLM conversations focus on language models and generated outputs. Production LLM work separates classification from generation. Teams compare hosted APIs with open-source models. Fine-tuning, retrieval, and vector databases become implementation choices. Latency, cost, and evaluation matter in production.^[5]

For narrower topics, use LLM production patterns, vector databases, and evaluation.

Generative AI covers systems that produce generated outputs such as text, code, summaries, and answers. The same bucket also includes translation work, image output, and planning.

Chatbot failures show generated outputs need security controls. Large-scale chatbot probing and knowledge-base exfiltration checks are part of that work. Teams also add output validation, query analysis, and layered defenses.^[6]

Those failure modes connect prompt engineering to security rather than treating a prompt as only copywriting.

Retrieval and Knowledge Systems

Many AI products become useful only when the model can work with the right context. Search systems connect NLP and personalization to learning-to-rank and LLMs. RAG adds retrieval plus generation to reduce hallucinations. It also brings in chunking, embeddings, and citations. Prompt design and human-in-the-loop evaluation stay part of the system.^[7]

That makes search and embeddings central AI topics, and RAG covers the broader architecture. RAG systems also inherit latency, cost, and data-quality problems. Retrieval can be a tool inside an agent rather than the whole system.^[8]

Agents and Action-Oriented AI

AI agents extend AI systems from answer generation into planning and tool use. They also add memory and action.

Agentic systems use objectives, tools, memory, and stores. They also depend on context design, custom evaluations, and mocked tools. Checks focus on outcomes rather than exact paths.^[8]

Agent design can borrow from reinforcement learning, evolutionary algorithms, early NLP, and prompt optimization. Minimal task decomposition can come before manager-agent orchestration. Tool integration and monitoring wrap the agent system.^[9] Use Game AI to LLM Agents for that history bridge. For design details, use agent engineering and multi-agent systems.

Evaluation, Reliability, and Cost

AI systems need evaluation because the output is probabilistic, context-dependent, and tied to user trust.

Production AI engineering starts with data trust and pipeline tests, then moves to prompt engineering and prompt evaluation. Engineers also manage prompt compression, prompt caching, and token use. Latency and model efficiency become engineering concerns too. ^[10] A useful AI feature must survive production traffic, not only a demo.

LLM and RAG evaluation use gold-standard examples, output-driven metrics, and human judgment. The work separates classification metrics from generative metrics.^[5]

Search and RAG systems add retrieval-level checks as well as answer-level checks.^[7]

Use LLM Evaluation Workflows and Production Search Evaluation for narrower measurement topics. Use Data Quality and Observability and Data Observability for data reliability.

Infrastructure and Model Control

Infrastructure-heavy AI work asks who controls compute, data, and models. Teams also have to own latency, privacy, and operating cost. Cloud and on-prem choices affect privacy, control, and decentralization. Training across multiple machines adds GPU coordination. Teams have to handle PyTorch and NCCL communication bottlenecks, Kubernetes limits, SLURM, and open-source orchestration.^[11]

The same infrastructure concern complements model deployment. Open-source models and hosted APIs sit alongside fine-tuning and serving optimization. Teams still have to manage latency tradeoffs.^[5] For deeper implementation work, use AI Infrastructure, Machine Learning Infrastructure, and MLOps.

Governance, Security, and Human Oversight

AI needs governance when it affects people, regulated decisions, or private data. Safety-critical workflows need the same discipline. Material business outcomes do too. Teams hit the same boundary in AI for social good work. Conservation monitoring, nonprofit optimization, accessibility support, and policy systems can turn model outputs into field work or access to resources.^[12][13][14]

Responsible AI work centers trust and fairness, plus explainability, stakeholder collaboration, and compliance. The work runs from data-level fairness checks and PII handling to cross-functional governance and human-in-the-loop oversight.^[15]

For LLM systems, governance also has a security layer. Prompt injection and data exfiltration sit beside hallucinations, controls, ROI, and hybrid human review.^[6]

Those risks are why the AI topic links out to Responsible AI and Governance, AI Red Teaming, Privacy Engineering for ML, and Security.

Related Pages

• AI Engineering and AI Engineer Role cover the builder role and skill stack.
• Generative AI, LLMs, and Prompt Engineering cover model behavior and generated outputs.
• Retrieval-Augmented Generation and RAG cover context and retrieval.
• AI Agents, Agent Engineering, and Multi-Agent Systems cover action-oriented AI systems.
• Data Products, MLOps, and Production cover delivery and operations.
• Evaluation, LLM Evaluation Workflows, and Production Search Evaluation cover measurement.
• AI Infrastructure and Machine Learning Infrastructure cover compute, orchestration, and deployment choices.
• Responsible AI and Governance, Governance, Security, and AI Red Teaming cover risk controls.