AI Singapore · AIAP

The AIAP-FDE Field Guide (5.0)

A self-paced path from zero to AIAP-ready, and on to forward deployment.

I created the AI Apprenticeship Programme to take people who can already build and turn them into AI engineers who ship real systems. Selection is rigorous, and it should be: the programme puts you on real projects with real consequences. This guide is how you get ready for it on your own time, with free material, at eight hours a week.

This edition is an evolution of nine years of the programme. Since 2018, more than five hundred engineers have come through AIAP. The closing stage is new: it comes from watching what forward deployed engineers do across the industry and what the role takes. The FDE wave is the market catching up with the apprenticeship model. The path runs in two stages: zero to AIAP first, then the forward deployed direction after it.

The guide stands on four pillars. A good AI engineer needs strong software engineering, because the AI stack is a software stack. You need AI, machine learning, and statistics, because the tools change and the foundations do not. You need large language models, agents, and agentic coding, because that is what the work is now. And you need ethics and governance, because Singapore leads here and AIAP engineers are expected to know it.

Every unit ends with something on GitHub. The portfolio is the proof; certificates are not. And each run of units ends with a gate. If you cannot pass the gate, repeat the run. Nobody is served by arriving at the AIAP assessment unprepared.

If you already know a topic, skip it. If a unit takes you longer, take the longer road. The gates, not the calendar, decide when you move on.

AIAP apprentices gathered at The Big Board room at AI Singapore during a presentation — From the version 4.0 guide: The Big Board, the room at AI Singapore where AIAP discussions, presentations, and meals happen.

The four pillars

Software engineering

The AI stack is a software stack: Python, the shell, git, containers, APIs, and databases.

AI, machine learning, and statistics

The classical core that does not go stale when the tools change.

LLMs, agents, agentic coding

The new layer: language models, retrieval, agents, and coding with AI as a practice.

Ethics and governance

AI Verify, the Model AI Governance Framework, and the risks specific to LLM systems.

What it actually takes

A unit in this guide is a unit of work, not a calendar promise. The stated workloads are the providers' own published figures; the estimated hours include the unit's deliverable, at eight hours a week, with optional depth excluded.

Unit	The courses	Stated workload	Estimated hours	Estimated weeks
1. Python from zero, properly	CS50's Introduction to Programming with Python	edX lists 12 weeks at 6 to 18 hours per week	75 to 110	9 to 14
2. Python depth, data, and SQL	Kaggle Learn (Python, Pandas)	Kaggle: Python 5 hours, Pandas about 4 to 5 hours	30 to 40	4 to 5
3. Databases	15-445 Intro to Database Systems	CMU: a semester of lectures, no self-paced total; Qdrant Essentials 7 days; the family intros about 1 hour each	30 to 45	4 to 6
4. The AI stack: containers and APIs	Get started with containers	IBM: about 4 hours for the path; the rest is documentation and building	25 to 35	3 to 4
5. The web layer	CS50's Web Programming with Python and JavaScript	Harvard states no total; the course is nine projects	70 to 110	9 to 14
6. Classical machine learning and stats	Essence of Linear Algebra; StatQuest	3Blue1Brown: 16 videos; StatQuest Statistics Fundamentals: 62 videos	30 to 40	4 to 5
7. ML end to end	Machine Learning Crash Course	Google: about 15 hours across 25 lessons	35 to 45	4 to 6
8. Deep learning	Neural Networks, Deep Learning, and AI (playlist); Practical Deep Learning for Coders	fast.ai: 9 lessons of about 90 minutes; the StatQuest playlist states no total	80 to 110	10 to 14
9. LLM engineering	LLM Course	Hugging Face: 12 chapters, 1 week each at 6 to 8 hours per week	72 to 96	9 to 12
10. Agents	AI Agents Course	Hugging Face: 5 chapters at 3 to 4 hours each	35 to 50	4 to 6
11. Evaluation and security for LLM systems	Your AI Product Needs Evals	An essay and reference material; no stated hours	20 to 30	2 to 4
12. Agentic engineering and production	Anthropic Academy	Claude Code in Action: 1 hour over 9 lessons; other courses unstated	35 to 50	4 to 6
13. Ethics and governance	AI Verify Foundation	Framework reading; no stated hours	12 to 20	2 to 3
14. Data engineering	Data Engineering Zoomcamp	DataTalks.Club: nine weeks as a cohort; self-paced anytime	80 to 120	10 to 15
15. Deployment infrastructure	Terraform tutorials	Step-by-step tutorials; no stated hours	25 to 45	3 to 6
16. Operations and reliability	MLOps Zoomcamp	DataTalks.Club: nine weeks of material; self-paced	40 to 60	5 to 8
17. Owning it end to end	Shape Up	A free online book plus a real deployment	40 to 80	5 to 10
Zero to AIAP (units 1 to 13)			550 to 780	69 to 98
The FDE direction (units 14 to 17)			190 to 310	23 to 38

Zero to AIAP-ready is about 550 to 780 hours; most people need a year to two years at eight hours a week. The forward deployed direction adds about 190 to 310 hours, and an apprentice meets much of that work inside the programme. The gates, not the calendar, decide when you move on.

The path

Zero to AIAP

Python from zero, properly

Software engineering

Learn to write, run, and debug real Python. Use the shell and git from the first week; an engineer without version control is not yet an engineer. Use an AI assistant as a study companion, and hold yourself to one rule: you must be able to explain every line you accept.

The courseCS50's Introduction to Programming with Python · HarvardThe best free first Python course; rigorous and graded.
Optional: The Missing Semester of Your CS Education (MIT).

The deliverable: A small command-line tool, tested, on GitHub, with a README.

Python depth, data, and SQL

Software engineering

Object-oriented programming and data structures at working depth, then the language of data: SQL. Keep the AI companion rule from unit one.

The courseKaggle Learn (Python, Pandas) · Google KaggleShort hands-on micro-courses in the browser.
Optional: SQL tutorial (Mode); Getting Started With Testing in Python (Real Python).

The deliverable: A data-cleaning command-line tool over a real messy CSV, with tests.

Databases

Software engineering

Where the data lives. You can query SQL; now learn what runs underneath it: storage, indexes, joins, and transactions. Carnegie Mellon's 15-445 is the course, and its lectures are free. Then meet the other families, each with the job it does best: documents for flexible records, graphs for relationships, vectors for similarity. The vector database returns in unit nine, because retrieval lives there.

The course15-445 Intro to Database Systems · Carnegie Mellon UniversityThe database systems course: storage, indexes, joins, transactions, query optimisation. Watch the lectures free; the projects are optional depth.
Optional: PostgreSQL tutorial (PostgreSQL); Qdrant Essentials (Qdrant); GraphAcademy (Neo4j); CRUD Operations in MongoDB (MongoDB University).

The deliverable: One dataset you know well, stored three ways: a normalised relational schema with the indexes that make it fast, a document collection, and a vector index you query by similarity; write down which family fits your workload, and why.

The AI stack: containers and APIs

Software engineering

Learn to package and serve what you build. Work through IBM's container learning path, which teaches with Podman itself: images, builds, volumes, and a push to a real registry. Then FastAPI, because an AI service is an API before it is anything else. This is the unit where software stops living only on your laptop.

The courseGet started with containers · IBM DeveloperA structured, free path, and the hands-on module teaches with Podman: run, build, volumes, and a push to a real registry.
Optional: Podman documentation (Podman); Container Internals (Red Hat Developer); FastAPI tutorial (FastAPI).

The deliverable: Unit two's tool, wrapped as a containerised HTTP API anyone can run, with the image pushed to a public registry.

The web layer

Software engineering

The systems you will deploy face people through a browser. Learn the web stack end to end: HTML, CSS, JavaScript, a framework, and the database behind it. CS50W keeps the Python from unit one and adds Django, React, and SQL depth, with testing and security along the way.

The courseCS50's Web Programming with Python and JavaScript · HarvardFree on OpenCourseWare; nine projects from HTML and Django through JavaScript and React, ending in scalability and security.
Optional: Full Stack Open (University of Helsinki).

The deliverable: Unit four's tool behind a web application a non-engineer can use: accounts, a database, and a link you can hand to someone.

Classical machine learning and stats

AI, machine learning, and statistics

The non-negotiable math floor: linear algebra and the statistics that machine learning stands on, taught visually before formally.

The courseEssence of Linear Algebra · 3Blue1BrownThe visual intuition for vectors, matrices, and transformations.
The courseStatQuest · Josh StarmerStatistics and classical ML concepts, one clear video at a time.
Optional: An Introduction to Statistical Learning (Python edition) (James, Witten, Hastie, Tibshirani, Taylor).

The deliverable: Topic summaries of five StatQuest concepts, written in your own words.

ML end to end

AI, machine learning, and statistics

Work MLCC end to end, then compete: enter a Kaggle competition and beat the baseline honestly.

The courseMachine Learning Crash Course · GoogleThe full pass, exercises included.
Optional: Intro to Machine Learning and Intermediate Machine Learning (Google Kaggle).

The deliverable: A Kaggle competition submission above the public baseline, with your feature reasoning and what the model cannot do written down.

Deep learning

AI, machine learning, and statistics

StatQuest takes you from gradient descent to transformers, one clear video at a time. fast.ai has you training and deploying from lesson one, and Karpathy builds a language model from nothing.

The courseNeural Networks, Deep Learning, and AI (playlist) · StatQuest with Josh StarmerFree on YouTube: gradient descent, backpropagation, attention, transformers, and the PyTorch intros, one step at a time.
The coursePractical Deep Learning for Coders · fast.aiLessons one to seven cover training, deployment, NLP, and from-scratch models.
Optional: The StatQuest Illustrated Guide to Neural Networks and AI (optional book) (Josh Starmer); Neural Networks: Zero to Hero (Andrej Karpathy).

The deliverable: A deployed image classifier, and the nano-GPT from the videos, running in your own repository.

LLM engineering

LLMs, agents, agentic coding

How large language models are actually used: prompting, embeddings, and retrieval. Build a question-answering system over documents you care about, and measure whether it retrieves the right passages. The vector database from unit three does its real work here.

The courseLLM Course · Hugging FaceTransformers, fine-tuning, and the ecosystem, hands on.
Optional: LADP Essentials (AI Singapore).

The deliverable: A retrieval-augmented question-answering app over your own documents, with a short written eval of its retrieval quality.

Agents

LLMs, agents, agentic coding

Agents that complete real multi-step tasks, with a review gate you designed before you launched them. Learn the protocols the ecosystem speaks and how to evaluate what comes back.

The courseAI Agents Course · Hugging FaceBuild and certificate included, free.
Optional: AI Agents for Beginners (Microsoft); Model Context Protocol (Anthropic).

The deliverable: An agent that completes a real multi-step task you defined, with its review gate and failure log written up.

Evaluation and security for LLM systems

LLMs, agents, agentic coding

An LLM system you cannot measure is a demo. Learn the discipline the frontier labs run on: eval sets, error analysis, and reading traces until the failure modes have names. Then the security pass, because deployed systems get attacked: prompt injection first among the risks.

The courseYour AI Product Needs Evals · Hamel HusainThe method essay: build an eval set, read the traces, and do error analysis until the failures have names.
Optional: OWASP Top 10 for LLM Applications (OWASP).

The deliverable: An eval harness for unit nine or ten's system: an eval set you wrote, the error analysis, and a one-page threat model against the OWASP list.

Agentic engineering and production

LLMs, agents, agentic coding

Agentic coding as a practice: you direct the tools, the tools do not own the outcome. Then production discipline: containers, tests, and a CI check, applied to your own LLM project.

The courseAnthropic Academy · AnthropicClaude Code in Action and the agentic engineering courses.
Optional: Machine Learning Systems Design (CS329S notes) (Stanford); Designing Machine Learning Systems (reading) (Chip Huyen); AIxTech (AI Singapore).

The deliverable: Unit nine or ten's project, rebuilt production-grade: containerised, tested, CI, and a README an engineer respects.

Ethics and governance

The governance block, when you have systems of your own to govern. Singapore's frameworks, the wider international picture, and the risks specific to LLM systems.

The courseAI Verify Foundation · AI Verify FoundationSingapore's AI governance testing framework and toolkit; start with What is AI Verify.
Optional: Model AI Governance Framework (IMDA); Practical Data Ethics (fast.ai).

The deliverable: A one-page governance review of your own project: risks named, mitigations stated, framework sections referenced.

After AIAP: the forward deployed engineer

Palantir made the title famous, and the frontier labs are hiring for it in Singapore: the forward deployed engineer, the engineer who sits next to the customer and owns whether the software actually works. OpenAI's posting for the role says it plainly: own technical delivery across deployments, from first prototype to stable production. An AIAP apprentice has already done a version of this work. Six months on a real project with a real sponsor is forward deployment under another name.

What separates an AIAP graduate from the role is specific. Enterprise data plumbing. Infrastructure as code, and the operations that keep a deployment alive. And posture: the FDE stands in front of the customer alone and owns the outcome end to end. Units fourteen to seventeen close what a self-paced guide can close. The two honest gaps no guide closes are software engineering mileage and solo ownership; you earn those by shipping for real customers, inside the programme and after it.

Take this stage when you are in the programme or past it. Apprentices meet much of this material inside the project phase, and these units are the map around it. If you finished the gates on your own, the work still stands: build for someone real, deploy it properly, and keep it running.

Data engineering

Real projects begin with plumbing. Extract messy data from where it lives, move it, transform it, and test what lands: orchestration, a warehouse, transformations with tests, then batch with Spark and streaming with Kafka.

The courseData Engineering Zoomcamp · DataTalks.ClubFree and open-source: Docker, Terraform, orchestration, BigQuery, dbt, Spark, and Kafka, ending in a capstone pipeline. It teaches on Google Cloud; the patterns travel.

The deliverable: A pipeline from a real messy source into a warehouse, with tested transformations and a run you can replay from nothing.

Deployment infrastructure

Stop configuring by hand. Infrastructure as code puts the whole stack in files you can review, version, and replay, and orchestration carries your containers onto clusters the way customers actually run them.

The courseTerraform tutorials · HashiCorpFree step-by-step tutorials, from the first configuration to provisioning and destroying real infrastructure.
Optional: Kubernetes tutorials (Kubernetes).

The deliverable: Your project's stack defined as code: tear it down to nothing and bring it back from the configuration alone.

Operations and reliability

Deployment is day one. The work is keeping the system alive: monitoring, alerting, safe rollouts, and an honest review every time something breaks.

The courseMLOps Zoomcamp · DataTalks.ClubFree and open-source: experiment tracking, pipelines, deployment, monitoring with Prometheus, Grafana, and Evidently, and continuous delivery. Take the deployment, monitoring, and best-practice modules.
Optional: Site Reliability Engineering (Google).

The deliverable: A system you deployed, monitored for real: dashboards, one alert that fired, and a written incident review of a failure you found.

Owning it end to end

The FDE stands in front of the customer alone. Scope the work with them, ship the smallest version that helps, and change it on real feedback. Shape Up carries the discipline: fixed time, variable scope, and the honesty to stop when the time is up.

The courseShape Up · 37signalsFree online: shaping work, betting on it, and shipping inside fixed time. Written for product teams; the method maps directly onto customer deployments.

The deliverable: One real user, served: the scope agreed in writing, the smallest useful version shipped, and a record of what changed because of their feedback.

The gates

Each run of units ends with a gate. If you cannot pass it, repeat the run. Nobody is served by arriving at the AIAP assessment unprepared.

Gate one: you can build software

Write, test, and debug a few hundred lines of Python without an AI assistant.
Explain every line of a program you wrote with one.
Use git fluently and query a SQL database.
Serve a small API in a container, with a web interface in front of it.
Say which database family fits a workload, and why.

Gate two: you can do machine learning honestly

Given a tabular dataset, produce a defensible model: correct splits, baselines, metric reasoning.
State what the model cannot do, in writing.
Explain bias and variance, overfitting, regularisation, and cross-validation without notes.

Gate three: you can build with LLMs

Explain attention and the transformer at the level of the Zero to Hero videos.
Build a retrieval system and an agent with a review gate.
Name the failure modes: hallucination, injection, retrieval miss; and say how you tested for each.
Build an eval set for your system, and report what it gets right and what it gets wrong.

Gate four: you are AIAP-ready

Solve mid-level Python problems fluently: the technical assessment's floor.
Show two portfolio artifacts on GitHub: one classical ML, one LLM or agent, each with production and governance discipline.
Whiteboard gradient descent, a decision tree split, cross-validation, attention, and why an LLM hallucinates.
Govern your own system: name its risks against Singapore's frameworks and what you did about each.
Know agentic coding tools' failure modes from experience, not theory.

Gate five: you can own it end to end

Provision a deployed system with infrastructure as code, tear it down, and bring it back from the code alone.
Keep a system alive: monitoring that shows, an alert that fired, and a written incident review.
Scope a piece of work with a real user, ship it, and change it on their feedback.
Write the eval set and the threat model for the system you deploy.

Selection

The AIAP technical assessment

Selection starts with the two-stage AIAP Proficiency Test. Stage one is a six-day take-home in Python: exploratory data analysis, and an end-to-end machine learning pipeline on an unseen dataset, built with software engineering rigour. Stage two is a technical interview followed by a collaborative group case study.

You need foundational Python, machine learning, and software engineering to pass it: units one through eight of this guide and the gate criteria are the preparation. Units nine through thirteen prepare you for the programme itself, where large language models and machine learning operations are part of the training. Units fourteen to seventeen are for during and after the programme: the forward deployed direction.

The full details and the application live at aiap.sg.