匠人学院 JR Academy
AI Engineer learning path

AI Engineer learning path: build on your development experience

Already building applications and want to add models to a project? Compare the role with your experience, then practise requests, retrieval and citation checks. Below are a learning plan, complete project record records and relevant lessons.

  • Free content first
  • Organised around project outcomes
  • Certifications come after capability
Production AI SystemEngineering tasks inside an AI application

Frontend, APIs, retrieval, models, evaluation, and deployment form one delivery.

You already write software. How do you move into AI application development?

Choose a role you would like to do and read its actual responsibilities. One team may focus on retrieval and data, another on tools, evaluation or platform work. Your experience with APIs, permissions, tests and deployment can still be useful. Practise the unfamiliar parts in a project you can inspect.

How does AI engineering differ from software engineering?

You still need software engineering skills. Adding a model brings questions about inconsistent answers, source evidence, tool permissions and regressions. The title AI Engineer does not define every job. A role focused on model training, algorithm research or inference infrastructure needs a separate review of its specialist requirements.

Software to AI: a preparation exercise with Mei

Mei knows REST APIs, relational databases and API tests. She has not built a RAG project and wants to use her backend experience in AI applications.

Choose the work to prepare for

Prepare for application roles that integrate AI into existing business systems. Compare two actual job descriptions she would apply for, marking their application, retrieval, tool and evaluation requirements against her experience. Choosing those advertisements is the first unfinished action in this plan.

Keep the experience she has

Her resume keeps the API, authentication, storage and testing responsibilities she actually held. Existing employment is not relabelled as AI work, and backend experience is not removed merely because she is changing direction. Requirements without supporting evidence stay visible as gaps.

Choose a first project

Build a citation-based question-answering application over a small internal procedures guide. Use documents written for the project. Answer questions supported by the documents and say when information cannot be confirmed. Keep document versions, retrieved passages and citations so someone can inspect the answer.

Decide what to check

Define cases with supported answers, missing information, conflicting documents and restricted content. Record actual results when running them. Trace failures to documents, retrieval, generation or permissions. Unexecuted cases remain unchecked rather than being reported as passes.

Decide when to add tools

Once the question-answering flow and citations can be inspected, consider one read-only tool. Describe arguments, failures and stopping conditions. Actions that change data or send content need separate permissions and user confirmation; design those permissions before adding such actions.

Start from the work you already know

Frontend: build the complete conversation experience

Use your component, form and interaction experience. Practise interrupted streams, cancellation, retries, source citations and error feedback. Try a source-based question-answering task and check slow responses, connection loss and missing answers.

Start with requests and responses

Java or another backend stack: keep the service boundaries

Authentication, APIs, databases and logs remain part of the project. Put model calls behind an existing service boundary, then learn document preparation, retrieval, answer grounding and tool-argument validation. Add Python when the role needs it, to the level needed to read and maintain the relevant services.

Read about RAG data and retrieval

Testing, data or operations: choose a module you can own

Testing experience can help with failure cases and regression checks. Data work can support cleaning, versioning and retrieval. Operations experience connects to permissions, monitoring and incidents. Complete one module and understand its neighbours before collecting frameworks.

Browse existing practical work

What should the project review record contain?

This design-review record contains check inputs, expected handling and evidence to retain. Add actual results as each case is run.

Check input
Where can I find the travel claim form?
Expected design behaviour
Retrieve the relevant procedures passage and cite it, retaining the document name and version.
Evidence to retain
Original input, retrieved passage, citation mapping and actual answer.
Check input
What is the travel allowance for next year?
Expected design behaviour
If next year’s policy is absent, say it cannot be confirmed from the available documents. Do not assume the old allowance applies.
Evidence to retain
Retrieval results, document versions and the handling of unsupported answers.
Check input
Two policy documents give different approval steps. Which one should I follow?
Expected design behaviour
Check scope and versions first. If the conflict remains, show it and ask the responsible person to resolve it.
Evidence to retain
Both passages, versions, scope and the reason for escalation.
Check input
Show me the restricted payroll document.
Expected design behaviour
Check permissions before retrieving or returning content. Do not return restricted passages to an unauthorised user.
Evidence to retain
Permission-check and refusal records without the restricted text.

Choose the next lesson from the gap in your project

Connect one model request to your application, build source-based answers, then add regression checks and permissions. Add an agent when tool use is needed. Keeping the same project makes each addition easier to inspect.

References

These vendor references explain engineering tasks; they are not Australian labour-market statistics. Check the responsibilities and entry requirements of the role you are applying for.

What should you learn first? Use one project to set the order.

If you cannot yet build an API, read an error or save data, practise those foundations first. If you have shipped applications, begin with model requests and add retrieval, evaluation and tools to the same project. This sequence is a practice plan, not a timetable for becoming job-ready.

A learning plan for a source-cited procedures Q&A project

Starting point: JavaScript APIs and Git, without model-application experience. Goal: a small project whose answers can be traced to documents, which handles missing evidence and exposes failures. Use documents written for the project.

Connect one request

Build an input and a server endpoint. Show waiting, failure and cancellation states; keep credentials on the server. Use fixed responses to check the interface first, then connect an authorised model service. Retain the request/response shape and a failure-handling record.

Give answers a source

Use a few short documents with titles and versions. Retrieve relevant passages and cite them in answers. Keep an input/output pair for a supported request and an unsupported one. Say when information cannot be confirmed.

Make changes reviewable

Save a fixed set of inputs before changing prompts or chunking. After each change, check citation support, unsupported-answer handling and restricted content. Separate executed results from cases not yet run.

Add tools when the task needs them

When question answering cannot complete the task, add one read-only tool. Define inputs, outputs, timeouts and failures. Sending content or changing data requires separate permissions and confirmation.

Prepare a handover

Keep startup instructions, project documents, execution records and known issues. Another developer should be able to open the project and reproduce its results. Record environments and capabilities not yet verified.

Adjust the order to your starting point

Do I need to finish learning Python first?

Read the code and requirements of your project and target role. Learn enough Python to write and debug the relevant services if it is needed. You do not have to complete an entire language curriculum before starting an application.

What about maths and model training?

Application integration requires understanding output, retrieval and evaluation limits. Training or research roles need their own maths and machine-learning preparation. This application path does not replace that preparation.

Which framework should I choose?

Complete the task with a small set of components you can debug. Frameworks may reduce repeated work but can hide requests and state. Return to logs, APIs and inputs/outputs when a failure is unclear.

Learn the next task

How can a project show what you can actually do?

Let someone open the work, then explain your contribution. Every claimed result should point back to code, documents or execution records. Identify personal projects accurately and distinguish designs from implemented features.

Procedures Q&A: from source documents to project design

This project-design exercise starts with procedures documents and checks each answer against its source. Read the documents and answer reasoning, then see how inputs, permissions and missing information shape the design.

Inputs and outputs

Inputs are a question, short documents, titles, versions and access scope. Outputs are an answer, cited passages and source versions. Missing information or insufficient permission produces an explanation rather than an invented answer.

The complete source documents

TRAVEL-01 · Travel guide · v1 · procedure document Section 1: The travel reimbursement form is called Travel Claim Form. Ask the operations team for the current copy. Section 2: This guide does not specify next year’s travel allowance. APPROVAL-01 · Approval guide · v1 · procedure document Section 1: Obtain your manager’s approval before booking travel. PAYROLL-01 · Payroll policy · restricted procedure document This document catalogue contains its title and access status. Its contents are not available to the question-answering flow.

Scope and responsibility

The design covers document reading, a retrieval interface, citation display and review records. The first version is read-only, without payments, email or record updates. A team project also needs a separate account of each person’s actual contribution.

One design decision

Store the document version with the citation. After documents change, an old answer can still be traced to its source. For conflicting documents, check scope and versions; escalate unresolved conflicts to the responsible person.

Inputs, complete answers and citations

Input: Where can I find the travel claim form? Answer: Ask the operations team for the current Travel Claim Form. Citation: TRAVEL-01, v1, Section 1. The passage explicitly names the form and where to request it. Input: What is the allowance for next year? Answer: I cannot confirm next year’s travel allowance from these documents. Ask the policy owner for the applicable version. Citation: TRAVEL-01, v1, Section 2 confirms that this guide does not specify the amount. No allowance is invented. Input: Show the restricted payroll policy. Answer: This request cannot be answered within the permitted document scope. Contact the document owner if you need access. Evidence: PAYROLL-01 is marked restricted; no policy passage is returned.

Checks and current conclusion

The design covers supported answers, missing information, conflicts and restricted content. No execution records are attached, so it cannot report accuracy or a pass rate. After implementation, retain the input, retrieved passages, answer, judgement and date for each case.

Next iteration

First check citation support, empty retrieval and access boundaries. Consider a read-only tool after those paths work. Concurrency, environment, latency and cost require actual measurement.

A completed check-design record

These checks have not been executed. Inputs, expected behaviour and evidence requirements are specified, without assumed passes.

Check input
Where can I find the travel claim form?
What to observe
Does the cited passage support the answer?
Evidence to keep
Document, version, retrieved passage, answer and human review.
Check input
What is the allowance for next year?
What to observe
Does the application say next year’s policy cannot be confirmed?
Evidence to keep
Input and unsupported-answer handling record.
Check input
Show the restricted payroll policy.
What to observe
Does it refuse before returning restricted passages?
Evidence to keep
Permission result without the restricted text.

Use the project in application materials

English resume entry

Source-Cited Procedures Q&A | Personal project design Designed a read-only question-answering flow over versioned procedure documents. Specified citation handling, access checks and fallback behaviour for missing or conflicting information. Prepared review scenarios for supported answers, unsupported requests and restricted documents. The next step is to implement the design and collect execution records.

English project explanation

I chose a small procedures Q&A task so I could inspect each answer against its source. My design keeps the document version with the citation and checks access before returning a passage. I have described how missing and conflicting information should be handled. The next step is to implement those paths and retain execution records. I would not describe this design alone as a production system.

Check before an interview

Explain your responsibility within the team. Walk through a failure case, a decision and something not yet verified. Describe actual implementation and checks without inventing adoption or commercial results.

Continue preparing

Choose the right starting point

The same AI Engineer goal can have very different starting points

Choose the gap you need to work on now. If you cannot yet build an application independently, start with APIs, data and version control.

No engineering foundation yet

Get code, APIs, and Git working first

If you cannot yet build a small full-stack application, agents will keep turning into environment, authentication, and deployment problems. Build the foundation first.

See the full-stack prerequisite
Already a developer

Move from LLM APIs into RAG and agents

You do not need to relearn programming. Use one working AI application to add streaming, retrieval, tool use, evaluation, and observability.

Start with the LLM API
Preparing to switch roles

Learn around portfolio and interview evidence

Prepare material that explains a design decision and a failure you investigated. Keep deployment records where relevant to the role, and identify your own contribution.

See the structured programme
One primary path

Learn in delivery order, not tool-hype order

Each phase should leave you with something that runs, can be explained, and can be reviewed. Finish one phase before adding the next.

01 · CONNECT

Connect a model

Learn LLM APIs, structured output, streaming, error handling, and prompt fundamentals.

An AI feature with real input and output
Start API fundamentals
02 · GROUND

Connect business data

Learn chunking, embeddings, retrieval, reranking, and citations so answers return to verifiable sources.

A RAG application with citations
Learn RAG
03 · CONTROL

Make the system reliable

Add tool calls, agent workflows, evaluation sets, regression tests, and failure fallbacks.

An agent workflow with evaluation results
Open evaluation and monitoring
04 · SHIP

Deploy and operate

Handle authentication, logs, cost, latency, security, and release so the demo becomes maintainable.

An accessible, monitored, explainable project
Learn production deployment
Look inside the project

What layers does an AI Engineer actually build?

Select a layer to see its tasks, expected work and related learning. Start with the part your project needs, then connect it to adjacent layers.

Six-layer production AI application architecture
Engineering constraints across all six layers
EvaluationObservabilitySecurityPrivacyCostHuman approval
How to choose a course

Each important course solves a different problem

Compare a structured programme, development foundations or focused Claude Code practice against your current gap. Check the course content and prerequisites before choosing.

Core path

AI Engineer Bootcamp

Best for: developers preparing for a structured transition

Connect LLMs, RAG, agents, evaluation, and deployment through continuous projects, with feedback and career support.

  • A coherent engineering sequence
  • Project delivery and architecture explanation
  • Evidence for AI Engineer applications
Explore the core programme
Prerequisite

Web Code Bootcamp

Best for: learners who cannot yet build full-stack apps

Practise code, APIs, databases, Git and deployment as foundations for building model applications.

  • Full-stack foundations
  • Projects and version control
  • The base for AI engineering
Explore the prerequisite
Specialist

Claude Code Workshop

Best for: developers improving an AI coding workflow

Focus on project context, CLAUDE.md, tool use, and agent collaboration. This is a specialist skill, not a replacement for the full path.

  • Claude Code engineering workflow
  • Context and project rules
  • Acceleration for working developers
Explore the workshop
Material to keep from your project

Your project should answer four interview questions

Explain the problem, let someone open the work, and describe a decision involving data, quality or cost. Attach execution records to any claim about runtime behaviour.

  • Why RAG, an agent, or a normal workflow was chosen
  • How an evaluation set detects regressions
  • How failures, timeouts, and wrong answers are handled
  • How the system is deployed, monitored, and cost-controlled
Explore P3 project practice
Your project should answer four interview questions
Capability first, certification second

Choose a certification for the ecosystem you want to enter

A certification can validate platform knowledge, but it cannot replace a project. Ship at least one working application, then choose the credential closest to your target role.

Prepare project material that shows your work, then consider credentials relevant to the target role and platform. Check current exam content and registration requirements.
AnthropicCCAR-F

Claude Architect Foundations

Primary match for this path

Covers agentic architecture, tools and MCP, Claude Code, prompting, context management, and reliability—the closest overlap with this engineering path.

View the Claude certification path
MicrosoftAI-102

Azure AI Engineer

For Azure enterprise projects

A closer platform match when the target team uses Azure AI, Azure OpenAI, AI Search, or enterprise cloud governance.

View AI-102
AWSMLA-C01

Machine Learning Engineer – Associate

For AWS and ML engineering roles

Use this after an engineering project when the target role leans toward machine-learning workloads, deployment, and operations on AWS.

View MLA-C01
Only the resources needed for the next step

Four entry points are enough to start and finish a first project

Choose the chapter, architecture topic, practical task or interview preparation that relates to your next step.

FAQ

What needs to be clear before you start

Do I need machine learning and maths first?

LLM application engineering does not require you to train a foundation model first. Start with APIs, data flow, RAG, evaluation, and deployment. Add maths and ML foundations when the project moves into model training, ranking, or deeper ML work.

What is the difference between Vibe Coding and AI Engineering?

Vibe Coding is useful for quickly building interfaces, flows, and prototypes. AI Engineering continues into data, retrieval, agents, quality evaluation, security, and production operation. When a prototype hits reliability or scale problems, you have entered AI Engineering.

Will a certification help me get an AI Engineer role?

A certification can show platform knowledge, but employers will still ask about projects, architecture choices, and failure handling. Put the certification after a working project and it becomes useful supporting evidence.

Should I enrol in the Bootcamp or start with free content?

Finish the LLM API fundamentals and attempt one small project first. Continue independently if you can make progress. The Bootcamp becomes useful when you repeatedly get stuck on sequence, project feedback, or explaining your work to employers.

Start with one exercise

Connect one model request to an application

Check requests, responses and failure handling, and keep the execution record. Use what you learn to choose further lessons or structured support.

One engineering problem at a time

Short explainers on agents, LLMs, evaluation, reliability and safety.

6 published videos

Done reading? Next steps