← SHANNONWAREARTICLE · 2026.08.29

AI Engineering Self Study

I was thinking about a way forward in my career, and asking Claude about the same. In mid June, I stumbled upon the AI / ML Engineering programme at SETU: 30 credit hours over one year conferring a QQI level 8 degree. Inquiring with the programme administration, I was informed that even though the application deadline was mid July, the programme was already massively oversubscribed, that the application cut off date had been moved up to end of June, and that my chances to gain admission were slim due to those reasons alone. Convinced that this is what I want to do most of all -- my "first choice" as it were, I made the application anyway and hoped for the best.

As a plan B, I fed the AI / ML Engineering course description into Claude, and asked it to create a syllabus for me, that I could follow myself, in the likely event that I did but get into the SETU programme.

Yesterday I found out that I in fact did not get into the AI / ML programme, but was offered to apply to a similar "computer science" course in the same track offered by SETU. I have decided to go ahead with my own plan B. Since I already have my own OpenClaw agent running in the cloud, I am keen to realise some practical returns from employing an AI agent. And given the pace at which time flies, if I don't find in-field employment before Spring 2027, I will re-apply to the AI / ML Engineering programme as soon as applications open.

Below is the course of study that Claude generated for me.

* * *

Claude:

A Self-Directed Curriculum in Artificial Intelligence and Machine Learning Engineering

Prepared for: Shannon Ware Structured as a parallel to: SETU Certificate in Artificial Intelligence and Machine Learning Engineering (Level 8, 30 ECTS) Format: Open courseware only — all core materials free (video, notes, code) Suggested cadence: ~10 hours per week over 6 months (extensible to 9–12 months at a lighter pace)

* * *

1. Purpose and design principles

This programme reconstructs the four SETU modules — Data Handling and Infrastructure for AI, The Evolution of Machine Learning and AI, Building Machine Learning Models for AI Systems, and the Applied AI Project — using entirely open, freely available courseware. It is deliberately not a beginner's path. It assumes and builds upon your existing competencies:

  • Programming: Fluent Python (3.13/3.14), C#/.NET, JavaScript/TypeScript, SQL.
  • Data: Prior exploratory data analysis with Pandas and Matplotlib (e.g. your nz-industry project); SQL query analysis at Fujitsu.
  • Cloud & infrastructure: Azure fundamentals (AZ-900, DP-900), Docker, Git/GitHub, CI/CD, a self-administered Linux VPS cluster, Flask-to-Heroku deployment (e.g. goats-path-woollen-mills).
  • Applied AI exposure: Hugging Face and OpenAI tooling; TTS work in terzo-audio, Ventriloquist, and StageMaster.

Because of this, the curriculum compresses the introductory material, moves quickly to rigour and engineering, and treats your own VPS cluster as the deployment target throughout — turning your infrastructure background into a genuine advantage rather than a subject to be learned from zero.

Three principles govern the design:

  1. Engineering-forward. The SETU programme is explicitly an engineering certificate assessed by project work, not a theory course. Every phase ends in something built, versioned, and (where possible) deployed.
  2. Portfolio as evidence. Each phase produces a public GitHub artefact. By the end you hold a coherent portfolio demonstrating the full lifecycle — the same evidentiary logic you applied to your RPL submission.
  3. Ethics and evaluation are not appendices. Responsible-AI and model-evaluation themes are woven through the modelling and deployment phases, mirroring the SETU emphasis on bias detection, robustness, and trustworthy systems.
* * *

2. Module mapping at a glance

| SETU module | ECTS | Curriculum phase(s) | Primary open resources | |---|---|---|---| | Data Handling and Infrastructure for AI | 10 | Phase 3 | Data Engineering Zoomcamp; MLOps Zoomcamp | | The Evolution of Machine Learning and AI | 5 | Phases 1 & 4 | MIT 6.S191; fast.ai; HF LLM Course; HF Diffusion Course | | Building Machine Learning Models for AI Systems | 5 | Phase 2 | CS229; Google MLCC; Made With ML; CS231n | | Applied AI Project | 10 | Phase 5 | Made With ML (end-to-end); ML Zoomcamp deployment |

* * *

3. Phase-by-phase curriculum

Phase 0 — Orientation and toolchain (Week 1)

A short calibration week, kept deliberately light given your background.

  • Stand up a clean working environment: a dedicated Python 3.10+ virtualenv (Made With ML pins 3.10), PyTorch, JupyterLab, and a GPU strategy. Since local training is discouraged unless you are comfortable with CUDA drivers, decide now between free Kaggle/Colab notebooks for training and your VPS for serving.
  • Create a single umbrella GitHub repository (e.g. ai-ml-engineering-portfolio) with a phase-per-directory structure and a running learning log — consistent with your "learning in public" habit.
  • Deliverable: A reproducible environment README and an empty-but-structured portfolio repo.

Phase 1 — Foundations and the evolution of ML/AI (Weeks 2–6)

Maps to: The Evolution of Machine Learning and AI (part 1)

The goal is a firm, code-first intuition for the arc from classical ML through deep learning to transformers and generative models — the exact conceptual sweep of the SETU "Evolution" module.

  • fast.ai — Practical Deep Learning for Coders (2022), Lessons 1–5. Work in notebooks; build and train a real image classifier and a tabular model in the first two lessons, then dig beneath the abstractions. Read the accompanying free book chapters in parallel.
  • MIT 6.S191 — Lectures 1–4 (Foundations, RNNs/sequence models, CNNs, generative models). Use these as concise, current lecture-format reinforcement of the fast.ai practical work. Complete the accompanying labs.
  • Google ML Crash Course — foundational modules (linear/logistic regression, generalisation, neural networks) as targeted gap-filling only where a concept feels shaky.
  • Deliverable: Two small trained models (one vision, one tabular) with a short written reflection tracing where each sits on the historical arc from symbolic AI to deep learning.

Phase 2 — Building and evaluating ML models rigorously (Weeks 7–11)

Maps to: Building Machine Learning Models for AI Systems

This phase supplies the rigour the SETU module demands: reliable, explainable models, proper metrics, and the failure modes — overfitting, bias, drift, adversarial fragility.

  • Stanford CS229 — selected lectures/notes on bias–variance, regularisation, evaluation, and learning theory. Use the open GitHub notes; you need the reasoning, not the full problem sets.
  • fast.ai — Lessons 6–7 (regularisation, collaborative filtering, and the training-loop internals such as SGD).
  • Made With ML — Model and Test lessons (training, tracking, tuning, evaluation; testing code, data, and models). This is where you learn to evaluate like an engineer, not just report accuracy.
  • Google MLCC — Fairness module and a self-directed reading on bias detection and robustness/drift, then reproduce a simple drift-detection experiment.
  • Deliverable: One well-evaluated model with a model card documenting metrics, a bias audit, an overfitting analysis, and a basic robustness/drift check.

Phase 3 — Data handling and infrastructure / MLOps (Weeks 12–18)

Maps to: Data Handling and Infrastructure for AI (10 ECTS — the largest module)

Your engineering strength should let you move confidently here. The aim is production data pipelines and the operational scaffolding around models.

  • Data Engineering Zoomcamp — core modules on containerised pipelines, workflow orchestration, the data warehouse, and batch vs. streaming (Docker, an orchestrator, a warehouse, Spark, and Kafka concepts). Skim topics you already command from Fujitsu/Terzo work; build the pipeline hands-on.
  • MLOps Zoomcamp — full arc: experiment tracking (MLflow), pipelines/orchestration, model deployment, and monitoring (Prometheus/Grafana), plus CI/CD with GitHub Actions. Prior Python, Docker basics, and the ML workflow are assumed — all of which you hold.
  • Deliverable: A batch-or-streaming data pipeline feeding a tracked, versioned model, with a monitoring dashboard — deployed to your own VPS to demonstrate infrastructure ownership.

Phase 4 — Deep learning, transformers, and generative AI (Weeks 19–22)

Maps to: The Evolution of Machine Learning and AI (part 2)

Now to the modern frontier the SETU module foregrounds: transformers, diffusion, LLMs, and responsible generative AI. Hugging Face recommends this material be taken after an introductory deep-learning course — which Phase 1 satisfies.

  • Hugging Face LLM Course — Chapters 1–9, progressing from Transformer internals and the Hub through fine-tuning to building and sharing demos (Gradio/Spaces). Chapters are designed at roughly 6–8 hours each; compress where the NLP basics overlap with your translation-domain intuition.
  • Hugging Face Diffusion Models Course — Units 1–2 for image/audio generation with Diffusers, connecting naturally to your existing audio/TTS project work.
  • Stanford CS224N — selected lectures on attention and transformers for deeper theoretical grounding, taken as needed rather than in full.
  • Deliverable: A fine-tuned transformer for a task of your choosing (text classification is a natural fit given your translation background), published to the Hub with a Gradio demo.

Phase 5 — Applied AI capstone (Weeks 23–26+)

Maps to: the Applied AI Project (10 ECTS)

An end-to-end, deployed AI solution — the definition of the SETU capstone: define the problem, collect and preprocess data, select/train/fine-tune, validate, and deploy via APIs and containerisation.

  • Framework: Follow Made With ML end-to-end (design → data → model → develop → test → reproduce → production), which is explicitly built to take you from prototype to a CI/CD-integrated, monitored production system.
  • Deployment craft: Draw the serving patterns from ML Zoomcamp — FastAPI, Docker, and optional Kubernetes/serverless — targeting your VPS cluster.
  • Project seed (choose one):
    • Extend terzo-audio into an ML-driven service (e.g. a fine-tuned model for text segmentation, quality scoring, or voice selection), giving your existing TTS utility a genuine learned component.
    • Extend nz-industry from descriptive EDA into a forecasting or classification system with a served API and a monitored pipeline.
    • A West Cork / maritime domain project leveraging your SNG context (e.g. equipment-inspection image classification or document triage), which would double as a distinctive portfolio piece.
  • Deliverable: A deployed, containerised, monitored AI application with a public repository, a technical write-up, and a short recorded demo — a portfolio artefact of the same standard the funded programme would require.
* * *

4. Weekly rhythm and cadence

A sustainable ~10-hour week, adaptable to your workload at SNG, Terzo Technical, and Goats Path:

  • ~5 hours — lectures/reading (the "what" and "why").
  • ~4 hours — hands-on coding, notebooks, and homework (the "how").
  • ~1 hour — writing: your learning log, model cards, and READMEs (the evidence).

If six months proves compressed alongside your commitments, stretch Phases 3–5 across additional weeks rather than reducing the hands-on share; the build-and-deploy work is where the engineering competence is actually forged.

* * *

5. Contingency and continuity with the funded route

You are pursuing the SETU place in parallel, and it remains possible you will be offered one. This curriculum is designed to complement rather than compete with that outcome:

  • If you are admitted, Phases 1–3 will have front-loaded the foundations, letting you engage the taught modules with unusual depth and freeing capacity for the assessed projects.
  • If you are not admitted this cycle, you will still hold a lifecycle-complete portfolio — foundations, rigorous modelling, a production pipeline, a fine-tuned transformer, and a deployed capstone — that stands on its own as demonstrable AI/ML engineering capability, and that strengthens any future RPL or fee-paying application.

Either way, the work is not wasted; it is the same competence the certificate is designed to certify.

* * *

6. Consolidated resource index

Foundations & deep learning

  • fast.ai — Practical Deep Learning for Coders (2022): https://course.fast.ai/ (free book: https://github.com/fastai/fastbook)
  • MIT 6.S191 — Introduction to Deep Learning: https://introtodeeplearning.com/ (labs: https://github.com/MITDeepLearning/introtodeeplearning)
  • Google Machine Learning Crash Course: https://developers.google.com/machine-learning/crash-course

Model building & evaluation

  • Stanford CS229 — Machine Learning (open notes): https://github.com/maxim5/cs229-2018-autumn
  • Stanford CS231n — Deep Learning for Computer Vision: https://cs231n.github.io/

Data engineering & MLOps

  • Data Engineering Zoomcamp: https://github.com/DataTalksClub/data-engineering-zoomcamp
  • MLOps Zoomcamp: https://github.com/DataTalksClub/mlops-zoomcamp
  • Made With ML (MLOps): https://madewithml.com/ (repo: https://github.com/GokuMohandas/Made-With-ML)
  • Machine Learning Zoomcamp (deployment): https://github.com/DataTalksClub/machine-learning-zoomcamp

Transformers, LLMs & generative AI

  • Hugging Face LLM Course: https://huggingface.co/learn/llm-course/chapter1/1
  • Hugging Face Diffusion Models Course: https://huggingface.co/learn (Diffusion course)
  • Stanford CS224N — NLP with Deep Learning: https://web.stanford.edu/class/cs224n/

All resources verified as openly available as of July 2026. Course cohorts and page structures evolve; where a live cohort has ended, the materials remain available for self-paced study.

ImpressumContactTermsPrivacy