What Alloy teaches

AI safety and governance is the discipline of making AI systems behave reliably, stay aligned with their operators’ intent, and remain accountable as they are deployed into real organisations. It spans the technical side — alignment techniques, guardrails, refusal and jailbreak resistance, evaluation of harmful or biased behaviour — and the organisational side — risk frameworks, model cards, audit trails, and the emerging regulatory landscape (the EU AI Act, NIST’s AI RMF, sector rules). As models take on higher-stakes work, the people who can reason about both the engineering and the policy of safe deployment become essential.

Multimodal AI engineering is building systems that reason over more than text — images, documents, audio, and video — and combine those signals to act. Modern models accept images and PDFs natively, which unlocks document understanding, visual question answering, OCR-free extraction, and screen/UI agents, but it also changes how you design prompts, manage tokens, and ground outputs. This track is about the engineering reality of multimodal pipelines: getting the inputs in cleanly, controlling cost and latency, and validating that the model actually grounded its answer in the pixels rather than hallucinating.

A retrieval-augmented system is only as good as the data pipeline behind it, and that pipeline is a data-engineering problem long before it is a model problem. This field covers the unglamorous work that decides RAG quality: ingesting and cleaning heterogeneous sources, chunking documents so meaning survives the split, choosing and tuning an embedding model, building and maintaining a vector index, and re-indexing as the corpus changes. It also covers the freshness discipline — knowing when stored knowledge has gone stale and re-validating it — which is exactly the problem Alloy’s own freshness pipeline is built to solve.

Fine-tuning and model adaptation is the art of deciding when a base model is not enough and changing its behaviour with your own data. It covers the full spectrum of adaptation techniques — parameter-efficient methods like LoRA and QLoRA, full supervised fine-tuning, and preference methods like RLHF and DPO — alongside the harder judgement of when adaptation actually beats better prompting or retrieval. The most valuable skill here is knowing the trade-offs: dataset curation, evaluation against a held-out set, the cost and serving implications, and catastrophic forgetting.

AI security and red-teaming is offensive and defensive practice for LLM-powered systems — adversarially probing them for the failure modes that classic application security does not catch. It centres on the attack surface unique to language models: prompt injection (direct and indirect, including data-exfiltration via tool calls), jailbreaks, training-data and system-prompt extraction, and the new risks introduced when agents are given tools and autonomy. The defensive side is just as concrete: input/output filtering, privilege separation for tools, sandboxing, and the monitoring that detects an attack in progress. This is among the fastest-moving security fields, because every new agent capability opens a new attack path.

LLMOps is the operational discipline of running language-model systems in production — the layer between a working prototype and a service real users depend on. It covers observability and tracing for non-deterministic systems, evaluation pipelines that run in CI so a prompt change cannot silently regress, cost and latency management, prompt and model version control, and safe rollout strategies (canaries, fallbacks, and rate limits). As organisations move from one-off demos to fleets of agents, the people who can keep those systems reliable, cheap, and observable become the difference between a pilot and a product.

Building with the Claude API is the base layer for almost everything else you build on Claude. It covers the Messages API end to end — system prompts, multi-turn conversations, streaming responses, controlling output with stop sequences and structured/JSON output, and the token, temperature, and cost levers that shape every call. It also covers the building blocks the higher tracks assume you already know: tool use at the API level, prompt evaluation, and the basics of retrieval and agent loops. Get this right and the agent, RAG, and eval tracks become a matter of composition rather than guesswork.

Claude Code Foundations is the daily-development workflow for building software with an agentic coding assistant — the discipline that separates "it wrote some code" from "it shipped a correct change." It covers how to give the model the right context (repo understanding, project instructions, file references), how to decompose a task so the agent stays on track, and the verification loop — running tests, reading failures, and iterating — that keeps generated code honest. It also covers the practical workflow: refactoring safely, debugging with the agent, and the project-instruction files that make every session start from your conventions.

The Model Context Protocol (MCP) is the emerging standard for connecting models to tools, data, and prompts through a uniform interface, and it is one of the fastest-moving integration layers in the field. This track covers building MCP servers and clients from the primitives up — tools, resources, and prompts — across the available transports, plus the security model (what an MCP server is allowed to do, and how to scope it), debugging with the Inspector, and how MCP plugs into Claude Code. It is the difference between bolting a one-off tool onto a model and exposing a reusable, governed capability any client can consume.

Commands, Skills, and Subagents are how a one-off Claude Code session becomes a reusable, team-shareable workflow. This track covers building custom slash commands that encode a repeatable task, packaging Skills the agent can invoke, and delegating to subagents so a complex job is split across focused workers with their own context. It also covers the operational details that make these reliable in practice — structured outputs, obstacle reporting, and tool limits — so a workflow behaves the same way every time instead of drifting. This is the layer that turns individual productivity into team leverage.

Tool use is what turns a language model from a text generator into something that can act — query a database, call an API, search the web, read files, or run code. This track covers the engineering reality of giving a model tools: designing tool schemas the model can call correctly, handling tool results and errors, scoping permissions so a tool cannot do more than intended, and the common built-in tools (web/search, files, code execution). It also covers when to integrate a tool directly versus through MCP, and the failure modes — malformed calls, hallucinated arguments, error loops — that separate a demo from a dependable integration.

The AWS Certified AI Practitioner is a foundational AI/ML/generative-AI credential with broad market demand and a 90-minute exam, aimed at people who need to speak credibly about AI on AWS without being deep ML engineers. This prep track maps the official exam guide to concrete study: the core AI/ML and generative-AI concepts, how the main AWS AI services fit together, responsible-AI and security basics, and the kind of scenario questions the exam favours. It pairs a readiness diagnostic with realistic practice sets so you walk in calibrated, not hopeful.

The AWS Certified Generative AI Developer – Professional is a developer-level credential for building generative-AI applications on AWS, a strong fit for serious cloud AI builders. This prep track works through the applied architecture the exam expects: building on Bedrock, application and RAG design patterns, governance and security for GenAI workloads, and alignment with the official practice questions. It is heavier and more hands-on than the practitioner track — the focus is on designing and defending real application architectures, not just recognising concepts.

The Google Cloud Generative AI Leader is a business- and strategy-facing certification — roughly 50–60 multiple-choice questions over 90 minutes, proctored — for people who lead or shape AI adoption rather than implement it. This prep track covers the generative-AI fundamentals the exam assumes, the Google Cloud AI offerings and where each fits, output-improvement techniques (prompting and grounding at a leadership level of detail), and the strategy and governance scenarios the exam favours. It is the non-engineering AI credential for decision-makers who still need to be technically literate.

Microsoft Azure AI Fundamentals is the entry-level Azure AI credential, currently mid-transition: the long-standing AI-900 is being retired and replaced by AI-901, so the smart move is to prep the durable fundamentals while tracking the new exam until it stabilises. This prep track covers the core AI, machine-learning, and generative-AI concepts on Azure that survive the version change, the main Azure AI services and where each fits, and responsible-AI principles Microsoft emphasises — built so the content shifts cleanly onto AI-901 once its guide is final rather than over-fitting a retiring exam.

The Databricks Certified Generative AI Engineer Associate is a proctored technical GenAI certification — 45 scored questions, 90 minutes — with real depth on RAG, vector search, and the Databricks platform. This prep track covers exactly that surface: retrieval and vector-search design, MLflow and model serving, governance and security on the Lakehouse, and the Databricks-specific architecture scenarios the exam tests. It is for engineers who want a vendor-backed, technically credible GenAI credential rather than a conceptual one.

The Salesforce Certified Agentforce Specialist is an enterprise agent/platform certification with strong CRM-driven demand and a concrete exam guide — 60 multiple-choice questions plus unscored items, 105 minutes, 73% to pass. This prep track maps that guide to study: Agentforce concepts and how AI agents are configured in Salesforce, the prompt-engineering and Data Cloud pieces the exam relies on, and the platform-configuration scenarios that make up the bulk of the questions. It is the credential for people building or administering AI agents inside the Salesforce ecosystem.

The NVIDIA-Certified Generative AI with LLMs (Associate) is a one-hour, associate-level exam (roughly 50–60 multiple-choice questions, two-year validity) carrying NVIDIA brand credibility, with a more fundamentals- and infrastructure-oriented slant than the application-focused cloud certs. This prep track covers the LLM and generative-AI fundamentals the exam tests, the model and infrastructure concepts NVIDIA emphasises, and the question style and pacing of a timed associate exam. It suits engineers who want a recognised, lower-cost entry credential anchored in the underlying technology rather than a single cloud vendor.

OpenAI has begun launching its own certification courses (AI Foundations, ChatGPT Foundations for educators, with a fuller certification path signalled), making it a very-high-priority watchlist credential — the right play is to build prep now against the public foundations material and adjust as official blueprints stabilise. This early prep track covers the OpenAI platform fundamentals, prompting and application patterns, and responsible-use material the announced courses center on, and is structured to convert quickly into a narrow, exam-specific track the moment OpenAI publishes a formal exam guide.

AI for Marketing is the applied track for marketers who want to use AI as a daily operating tool rather than a novelty — and to do it without producing generic, off-brand, or factually wrong output. It covers building a reusable brand-voice and context system the model works from, the content workflows that actually save time (briefs, drafts, repurposing one asset into many), AI-assisted research and audience analysis, and the judgment to keep a human in the loop where claims, compliance, and taste matter. The emphasis is on workflows you can trust and repeat, not one-off prompts.

AI for Sales is for sellers and sales teams who want AI to compress the busywork — research, personalisation, follow-up, CRM hygiene — without sounding robotic or losing the relationship. It covers using AI for fast, accurate account and prospect research, drafting personalised outreach that still reads human, summarising calls and keeping the CRM honest, and preparing for conversations with AI-built briefs. It also covers the line you should not cross: where AI assists and where the human must own the relationship, the claims, and the judgment.

AI for Customer Support is the applied track for support teams deploying AI on the front line — where a wrong or made-up answer directly damages trust. It covers grounding AI responses in real, current knowledge (so the assistant cites the help centre instead of confabulating), designing the human-in-the-loop and escalation paths that catch the cases AI should not handle alone, drafting and triaging at speed, and measuring quality honestly. The throughline is the same discipline the rest of Alloy teaches: AI that is grounded and accountable, not just fast.

AI for PR & Communications is for comms professionals who operate in high-stakes, reputation-sensitive territory where a sloppy or fabricated line is a real liability. It covers using AI to accelerate the safe parts — research, message drafting, repurposing across formats, monitoring and summarising coverage — while building the review discipline that protects accuracy, tone, and on-the-record claims. It also covers handling sensitive and crisis communications, where AI can prepare and structure but a human must own every published word.

AI for Research & Analysis is for analysts, knowledge workers, and anyone who turns large amounts of source material into a defensible conclusion. It covers using AI to gather, structure, and synthesise information across documents, the critical skill of verifying and citing what the model claims (rather than trusting a confident summary), and techniques for analysing data and surfacing patterns while staying alert to hallucination and bias. The core lesson is epistemic discipline: AI dramatically speeds up research, but the human owns whether the conclusion is actually supported.

AI for Founders is the operator track for early-stage builders who have to be marketing, sales, support, ops, and product at once — and want AI to act as a leverage multiplier across all of it. It covers using AI to move fast across functions without producing throwaway output, deciding where AI genuinely creates leverage versus where it wastes time, building lightweight reusable systems (context, prompts, workflows) instead of starting from scratch each time, and keeping a hand on quality and truth when there is no one else to check the work. It is breadth with judgment, tuned to the founder reality of doing everything.

A Skill Alloy is a durable, job-role composition — the specific blend of concepts and proven skills, drawn from multiple tracks, that a real role demands. The AI Engineer alloy binds the engineering spine into one coherent target: prompting and structured output, agents and tool use, RAG and memory, evaluation and reliability, and the production operations that keep it all running. Rather than "finish these courses," it answers "what must I be able to demonstrably do to be an AI engineer" — and the Alloy Path routes you through exactly the elements you are missing, tracked against your Skill Evidence Graph.

The RAG Engineer alloy is the job-role composition for people who build and operate retrieval-augmented systems end to end — where the work is as much data engineering as it is modelling. It binds the data pipeline (ingestion, chunking, embeddings, vector indexing, freshness), the retrieval and grounding skills that make answers faithful, and the evaluation discipline that proves retrieval quality rather than assuming it. Pursued as an Alloy, it routes you through only the elements you have not yet demonstrated and tracks each against your evidence graph, so "RAG Engineer" becomes provable, not just claimed.

The AI Product Manager alloy is the composition for people who ship AI products and need technical literacy plus product judgment — enough to scope what is feasible, evaluate quality honestly, and reason about cost, risk, and governance, without necessarily writing the model code. It blends the conceptual core (how models, prompting, RAG, agents, and evaluation actually work), the operator-level judgment of where AI creates value versus risk, and the governance and reliability awareness that keeps a launch defensible. As an Alloy it maps the role to demonstrable competencies and routes you through the gaps.

The AI Solutions Architect alloy is the composition for people who design end-to-end AI systems and have to make the high-leverage architectural calls — which pattern (prompting vs RAG vs fine-tuning vs agents), which integration layer (direct tools vs MCP), how to evaluate and secure it, and how to operate it at scale. It binds the cross-cutting skills that span the whole spine — architecture judgment, integration design, evaluation, security, and operations — into one role target. Pursued as an Alloy, it identifies which of those architectural competencies you have demonstrated and routes you through the rest.

How modern AI agents remember — the architecture of semantic, episodic, and procedural memory, and the retrieval pipelines that keep an agent's knowledge accurate over time. You leave able to design a memory layer that stays fresh instead of silently rotting.

What turns a language model into an agent — the loop that drives its own control flow, the coordinator that routes its decisions, and the tools and MCP servers that give it real capabilities. You leave able to reason about agent architecture and design tools the model uses correctly.

How to know whether an AI system is actually good — and keep it good. The discipline of evaluation, grounding, and the failure modes that make a confident answer quietly wrong. You leave able to design evals and reliability checks that catch silent failures before your users do.

How to put AI to work in a business without writing code — choosing where it actually pays off, designing a human-in-the-loop workflow you can trust, and measuring whether it earned its place. Interactive practice, hands-on labs, and the rigorous final for this track are coming as the concept graph for this domain is populated; the two lessons below are real and complete today.

Redesign how you work so AI removes the friction instead of adding a new tool to manage. Map your real workflow, find the AI-shaped bottlenecks, and wire assistants into the seams — using the kinds of tools Memriq builds (Helm for tasks, Harbor for knowledge capture, Haven for tracking). Interactive practice, labs, and the rigorous final are coming as the concept graph is populated; the two lessons below are real and complete today.

Why long-context agents degrade, and the engineering that stops them. You leave able to diagnose context rot, budget a window deliberately, compact without losing the thread, and ground answers so a model refuses instead of inventing.

How to drive a language model deliberately — the difference between asking nicely and actually engineering a prompt and its context window so the model returns output you can parse and act on every time. You leave able to design prompts and context that produce reliable, machine-readable results instead of prose you have to scrape.

Build a credible professional presence with AI as your analyst and editor, not your ghostwriter. Learn to find your real positioning, turn it into a consistent voice, and use AI to sharpen — never fake — what you publish. Interactive practice, labs, and the rigorous final for this track are coming as the concept graph is populated; the two lessons below are real and complete today.