Foundation
Weeks 1–4: individual foundations, project launch and a controlled baseline.
EAU OPEN LEARNING · EOL-AIA-001
Advanced technical · FREE PRACTICAL TRAINING
Advanced Applied AI: LLM Engineering, RAG & Intelligent Automation
Build Atlas from data engineering and ML baselines through structured LLM outputs, grounded RAG, evaluation and a human-approved workflow. Every claim is tied to a measurable test or a documented limitation.
PROGRAMME AT A GLANCE
Engineer an AI system that can be tested—not just demonstrated.
12 weeks · 24 live hours · approximately 24 guided practice hours. Workload estimates are not academic credits.
Who it is for
- Python practitioners
- Data or ML professionals
- Engineers building testable AI workflows
Entry readiness
- Working Python and notebook experience
- Foundational supervised ML and evaluation
- Comfort reading technical English documentation
Atlas
Atlas is a fictional bilingual service-workflow prototype that classifies synthetic tickets, retrieves approved knowledge, drafts grounded responses and requires human approval.
THREE CONNECTED PHASES
Each month raises the standard of evidence and delivery.
Build and test
Weeks 5–8: deeper application, documented decisions and staged evaluation.
Integrate and hand over
Weeks 9–12: integrate evidence, transfer the skill and deliver the final portfolio.
COMPLETE PUBLIC CURRICULUM
Twelve weeks. A clear practical deliverable every week.
Open any week to review the focus, practical work and public deliverable. Internal cases, answer keys and restricted data are not published.
W01AI Problem Framing & System Architecture
Focus: Turn a business request into an AI system specification, baseline, success measures and risk boundaries.
Practical work: Problem Canvas, baseline test on 20 synthetic cases, AI risk card.
Deliverable: `W01_Problem_and_Evaluation_Card.pdf`.
W02Data Engineering & Project Kickoff
Focus: Audit a synthetic bilingual ticket set, design the data contract, prevent grouped data leakage, freeze the split.
Practical work: Client scope, schema checks, duplicates, data dictionary, group-aware split, dataset card.
Deliverable: `W02_Data_Pipeline.ipynb + Data_Contract.md`.
W03Supervised ML Baseline & Error Analysis
Focus: Build a text classification baseline and interpret macro-F1, per-class recall and confusion matrix.
Practical work: TF-IDF + linear classifier, frozen holdout, error taxonomy, compare to rule baseline.
Deliverable: `W03_ML_Baseline.ipynb + Model_Evaluation.pdf`.
W04Responsible AI & Foundation Release Gate
Focus: Review privacy, language slices, dataset provenance and unsupported uses before accepting baseline.
Practical work: Model card, subgroup review, risk register, go/correct/hold decision.
Deliverable: `W04_Model_Card.pdf + Risk_Decision_Log.xlsx`.
W05Foundation Models & Structured Generation
Focus: Design an LLM output contract with schema validation, semantic checks, failure handling and abstention.
Practical work: Prompt versioning, bounded JSON schema, test invalid outputs and policy conflicts.
Deliverable: `W05_Structured_Output.ipynb + Validation_Results.csv`.
W06MLOps, Reproducibility & Observability
Focus: Convert ad hoc notebooks to a repeatable pipeline with tests, versioned configuration and privacy-aware traces.
Practical work: Clean-run, frozen requirements, timeout and retry, mock-provider fallback, redacted trace.
Deliverable: `W06_Reproducible_Pipeline.zip + Run_Manifest.json`.
W07RAG Engineering & Source Grounding
Focus: Build a retrieval-augmented draft assistant over approved knowledge, with version-aware citations and abstention.
Practical work: Chunking, retrieval baseline, source IDs, answerability, retrieval Recall@k.
Deliverable: `W07_RAG_System.ipynb + Retrieval_Evaluation.csv`.
W08Multimodal Intake & System V1 Evaluation
Focus: Handle one bounded scanned-document intake case and assess extraction uncertainty and end-to-end errors.
Practical work: Synthetic scan, parser/OCR with text fallback, uncertain extraction quarantine, system V1 scorecard.
Deliverable: `W08_Multimodal_Review.pdf + System_V1_Scorecard.xlsx`.
W09Intelligent Orchestration & Human Approval
Focus: Connect classification, retrieval and draft generation into a stateful workflow with mandatory human approval.
Practical work: State machine, tool allowlist, sensitive-case routing, failure-safe states, no external auto-send.
Deliverable: `W09_Atlas_Workflow.zip + Routing_Test_Results.csv`.
W10AI Evaluation, Experimentation & Unit Economics
Focus: Compare baseline versus new system on a frozen benchmark across correctness, support, abstention, latency and measured cost.
Practical work: Scorecard, per-slice errors, A/B experiment plan, actual vs estimated costs.
Deliverable: `W10_Evaluation_Report.pdf + Frozen_Scorecard.csv`.
W11Reliability, Security & Client Readiness
Focus: Test prompt injection, unsupported questions, expired sources and sensitive actions; package a reproducible sandbox release.
Practical work: Safety cases, risk and limitation register, README, installation steps and client handover.
Deliverable: `W11_Release_Candidate.zip + Safety_Results.csv`.
W12Independent Transfer Test & Capstone Handover
Focus: Solve an unseen individual case, then defend the team prototype, results and operating limits.
Practical work: Blind new case, source-aware decision, escalation, live or recorded demo, client acceptance pack.
Deliverable: `W12_Individual_Transfer_Test + Final_Atlas_Pack`.
YOUR PRACTICAL PORTFOLIO
What you will document through the pathway.
Tools: Python, notebooks, open-source ML tooling and an approved mock/provider fallback. No mandatory participant-paid API or GPU.
Professional boundary: A tested sandbox, not unattended production deployment, model training from scratch or guaranteed commercial readiness.
- Runnable sandbox and README
- Data contract and model card
- Evaluation scorecard and failure analysis
- RAG citation and abstention records
- Human review and safety workflow
USEFUL QUESTIONS
Understand the boundaries before you choose.
Is this suitable for complete beginners?
No. It assumes working Python and foundational ML knowledge so the live time can focus on engineering and evaluation.
Will the workflow send messages automatically?
No. The learning design requires explicit human approval and uses synthetic or authorised material.
EAU OPEN LEARNING
Review the details, then start your free application.
This is free independent non-degree training. It creates no academic student number and requires no payment. Moodle access remains a manual step after training acceptance.