Environment setup, Canvas orientation, and cohort welcome. Participants confirm the Python environment, API credentials, and access to all course tools before Module 1 begins.

Objective: Understand what LLMs are, how they process and generate text, and how to run inference tasks against them.

Content:

• Introduction to LLMs: Overview of how LLMs are transforming industries, including healthcare, finance, and e-commerce, with real-world examples of deployed applications across text summarisation, compliance, and personalisation
• What are Large Language Models? The “Core 4” foundational concepts: transformer architecture and attention mechanisms; tokenisation and model inputs; the model lifecycle (pre-training and fine-tuning); and scaling laws, emergent behaviours, and limitations
• Process of LLMs: How text is broken into tokens, processed through context windows, and used during inference; distinction between masked language modelling (BERT-style) and auto-regressive modelling (GPT-style), and how each affects output behaviour
• Example – Utilizing LLMs for real-world tasks: Applied walkthroughs across three industry scenarios covering clinical record summarisation, e-commerce recommendation parsing, and automated regulatory compliance
• Technical Lesson – LLMs in practice: Hands-on implementation of inference tasks using BERT and GPT-based models via Hugging Face Transformers, including parameter experimentation and output documentation

Daily Lab

• LLM Inference Tasks
  • Apply a BERT-based model (bert-base-uncased or distilbert-base-uncased) to predict masked tokens in code-related sentences, and use distilgpt2 to generate debugging suggestions from a text prompt
  • Adapt starter code from the Technical Lesson for new inputs, experimenting with at least three prompt structures (zero-shot, few-shot, parameter-adjusted)
  • Document how prompt structure and parameter changes affect output quality and submit the Jupyter Notebook with observations

Tools: Hugging Face Transformers, PyTorch, bert-base-uncased / distilbert-base-uncased, distilgpt2, Jupyter Notebook or Google Colab

Objective: Design and evaluate advanced prompt structures to optimize LLMs outputs for specific business tasks.

Content:

• Introduction to Advanced Prompt Engineering: Overview of how prompt design directly shapes LLM output quality, with industry examples across financial services (multi-turn client advisory), e-commerce (code assistance and API debugging), and healthcare (patient triage chatbots)
• What is Advanced Prompt Engineering? Core terminology and techniques: prompt structuring and design, few-shot and in-context learning, chain-of-thought reasoning, temperature settings, and iterative refinement
• Process – Advanced Prompt Engineering with LLMs: How to craft prompts with explicit context, embedded examples, and sequential reasoning cues; strategies for managing multi-turn conversations and maintaining context across successive interactions
• Example – Advanced Prompt Engineering with LLMs: Applied walkthroughs demonstrating how prompt variations affect output precision and relevance across code assistance, debugging, and domain-specific query tasks
• Technical Lesson – Advanced Prompt Engineering with LLMs: Hands-on implementation of prompt engineering techniques via the Hugging Face API, including few-shot prompt construction, chain-of-thought cues, and parameter adjustment for business-specific outputs

Daily Lab

• Advanced Prompt Engineering with LLMs
  • Retrieve and securely store a Hugging Face API key as an environment variable, then select one of three provided buggy code snippets (Pandas/sklearn, TensorFlow, or PyTorch) as the debugging target
  • Design a custom prompt incorporating in-context examples and chain-of-thought cues, execute it via the Hugging Face API, and evaluate output quality
  • Iterate on the prompt structure across at least two variations, documenting how changes affect response accuracy and relevance

Tools: Hugging Face Transformers API, Jupyter Notebook or Google Colab / Kaggle, Python (pandas, sklearn, tensorflow, or torch depending on chosen code snippet)

Objective: Apply fine-tuning techniques to adapt pre-trained models for domain-specific tasks using curated datasets.

Content:

• Introduction to Fine-Tuning Pre-Trained Models: Overview of how fine-tuning adapts generalised models for targeted applications, with industry examples across healthcare (clinical diagnostic assistant), finance (automated risk assessment), and e-commerce (customer sentiment analysis)
• What is Fine-Tuning? Core concepts covering transfer learning, fine-tuning strategies (hyperparameter adjustment, data augmentation), proprietary business model customisation, and evaluation and optimisation practices
• Process – Fine-Tuning Pre-Trained Models: Step-by-step workflow covering data preparation, model selection, hyperparameter configuration, iterative training, and performance evaluation post-fine-tuning
• Example – Fine-Tuning Pre-Trained Models: Applied walkthroughs demonstrating how domain-specific datasets transform generalised model outputs into accurate, context-aware tools for clinical, financial, and retail applications
• Technical Lesson – Fine-Tuning Pre-Trained Models: Hands-on implementation of transfer learning using Hugging Face Transformers, including fine-tuning a pre-trained model on a curated dataset and evaluating performance against the base model

Daily Lab

• Fine-Tuning Pre-Trained Models
  • Identify errors in the provided starter code for a logistics-themed fine-tuning pipeline, then use an LLM to debug them, documenting prompts used and evaluating which responses were most actionable
  • Apply corrections, re-run the fine-tuning pipeline, and test the updated model against simulated delivery scenario inputs
  • Compare fine-tuned model performance against the base model and submit a 2-3 paragraph report covering errors found, changes made, and reflections on using LLMs as a debugging aid

Tools: Hugging Face Transformers, PyTorch, Jupyter Notebook, or Google Colab

Objective: Manage the full ML lifecycle using the open-source MLOps stack, including experiment tracking, versioning, and monitoring.

Content

• Introduction to ML Operations: Overview of how MLOps converges machine learning with software engineering practices, with industry examples across personalisation at scale (Netflix), fraud detection (Airbnb), and predictive maintenance (Siemens)
• What is ML Operations? Core components of a production-grade MLOps pipeline: data intake and feature management, experiment tracking and training, and automated validation (CI); how DevOps principles reduce deployment time and establish model refinement feedback loops
• Process – ML Operations: End-to-end ML lifecycle workflow covering data ingestion, cleaning, and feature versioning; reproducible model training with logged hyperparameters and metrics using MLflow; and CI practices for testing data pipelines and validating model performance before deployment
• Example – ML Operations: Applied walkthroughs demonstrating how MLOps frameworks support model lineage, experiment comparison, and safe production deployment across real-world organisational contexts
• Technical Lesson – ML Operations: Hands-on implementation of an MLOps pipeline using sentence-transformers, scikit-learn, and mlflow, including experiment tracking, confusion matrix interpretation, and CI-style validation checks for an LLM-based ticket routing system

Daily Lab

• MLOps for LLM-Based Ticket Routing
  • Convert support ticket text into semantic embeddings using sentence-transformers (all-MiniLM-L6-v2) and train three classifiers (Logistic Regression, Random Forest, SVM) on the embedded data
  • Track all experiments using MLflow, logging model type, embedding model, accuracy, and F1 score, then generate and interpret confusion matrices for each model
  • Run CI-style validation checks on embedding dimensionality and class distribution, and answer interpretation questions on LLM embeddings vs. TF-IDF, MLflow reproducibility, and production validation risks

Tools: sentence-transformers, scikit-learn, mlflow, matplotlib, seaborn, Jupyter Notebook, or Google Colab

Objective: Deploy ML models to production environments and manage the full pipeline from MLOps to live systems.

Content:

• Introduction to ML Deployment: Overview of how ML Deployment takes validated models into production, with industry examples across dynamic pricing (Uber), fraud detection (Capital One), and music recommendation (Spotify), illustrating packaging, monitoring, and governance in practice
• What is ML Deployment? Core distinction between MLOps and ML Deployment, including key topics covering packaging and continuous delivery (CD), production monitoring and observability, and governance and continuous improvement
• Process – ML Deployment: End-to-end deployment workflow covering model packaging into portable artifacts, blue-green and canary deployment strategies, live monitoring of latency and prediction drift, fairness checks, audit logging, access controls, and automated retraining triggers
• Example – ML Deployment: Applied walkthroughs demonstrating how containerised models are deployed behind RESTful APIs, monitored with observability dashboards, and governed with versioning and rollback mechanisms across real-world organisational contexts
• Technical Lesson – ML Operations to Deployment (Parts 1 and 2): Hands-on implementation of a full deployment pipeline using Flask, pickle, and mlflow, including model packaging, REST API serving, prediction monitoring, and model card generation for governance and auditability

Daily Lab

• ML Deployment
  • Select the best-performing model from the Module 4 lab using F1 scores, package it into a versioned deployment directory with model weights, metadata, and example inputs, then serve it via a Flask REST API with /predict and /health endpoints
  • Simulate production monitoring by generating a prediction distribution plot and identifying signs of class drift or traffic imbalance
  • Generate a model_card.json documenting performance summary, ethical considerations, and retraining criteria, and answer interpretation questions on deployment packaging, drift detection, and governance

Note: This lab is a direct continuation of the Module 4 MLOps lab. Therefore, completion of this lab depends on completing the previous module.

Tools: Flask, pickle, mlflow, sentence-transformers, scikit-learn, matplotlib, Jupyter Notebook, or Google Colab

Objective: Build agentic AI systems using LangChain and the Model Context Protocol to solve real-world problems.

Content:

• Introduction to Agentic AI: Overview of how agentic AI systems autonomously perceive, reason, plan, and act toward specific goals, with industry examples across logistics (Dow invoice analysis), professional services (BDO Colombia process automation), and IT support (Power Design HelpBot)
• Defining Agentic AI: Core concepts covering agent classification (simple reflex, model-based, goal-based, utility-based, and learning agents); hierarchical vs. decentralised architectures; and how agentic systems differ from traditional and generative AI models
• Process – Agentic AI Frameworks: How LangChain enables tool use, memory, and planning within agent workflows; how the Model Context Protocol (MCP) standardises context sharing between agents and external systems; and the seven-step process for developing and deploying agentic AI solutions
• Example – Developing Agentic AI Solutions: Applied walkthroughs demonstrating how agentic frameworks automate complex, multi-step data workflows across real-world organisational contexts, from invoice discrepancy detection to natural language IT support
• Technical Lesson – Building a Data Analysis Agent: Hands-on implementation of a data analysis agent using LangChain and MCP, including tool integration, reasoning chain design, and deployment of a multi-step agent pipeline

Daily Lab

• Agentic AI for Transportation and Logistics
  • Define SMART objectives, operational constraints, and autonomy level for a route optimisation agent; map at least five data sources and identify required tools/APIs
  • Select and justify an agent architecture and classification type, designing at least three sub-agents with defined responsibilities and an MCP integration plan
  • Produce a workflow state diagram from shipping request to finalised route, identifying critical decision points, edge case handling, and an uncertainty management strategy
  • Propose a testing plan with scenarios and baseline comparisons, then outline a phased deployment plan with KPIs and human oversight processes

Note: This is a conceptual design exercise. No programming environment is required. All responses are submitted via the provided Google Doc lab template.

Tools: Google Docs (lab template provided), LangChain, and MCP concepts from the Technical Lesson as reference frameworks

Objective: Apply ethical frameworks to AI system design, addressing bias, governance, regulation, and responsible deployment.

Content:

• Introduction to AI Ethics: Overview of the ethical, legal, and societal implications of generative AI, with industry examples across copyright and IP (GitHub Copilot), privacy and consent (Clearview AI), and responsible innovation (Google DeepMind AlphaFold)
• What is AI Ethics? Core considerations covering legal landscapes and fair use, data ethics and informed consent, licensing and attribution, bias and fairness, and transparency and explainability in consequential AI systems
• Defining AI Regulation and Governance: Current legal frameworks governing AI development and deployment; how regulatory cases and settlements are reshaping industry practices around data collection, biometric privacy, and intellectual property
• Process – AI Ethics in Data Science: How ethical principles translate into concrete technical decisions: detecting and mitigating bias in training data, implementing fairness checks across demographic groups, designing for transparency and accountability, and integrating ethics review into the ML development lifecycle
• Technical Lesson – AI Ethics in Practice: Hands-on application of an ethics impact assessment framework to a real or hypothetical AI system, identifying bias risks, governance gaps, and recommended mitigations

Daily Lab

• AI Ethics Impact Assessment
  • Analyse the MediPredict health monitoring system, classify its overall risk level, identify all affected stakeholder groups, and assess ethical issues across all five data sources (privacy, consent, data quality, bias)
  • Complete an impact analysis table for at least five stakeholder groups with severity and likelihood ratings, then identify the 3-4 most critical ethical concerns
  • Develop mitigation strategies for each top concern, covering proposed approach, implementation method, measurable success criteria, and trade-offs
  • Design a governance framework and produce a final summary covering overall risk rating, key mitigations, development team recommendations, and pre-deployment conditions

Note: This is a written assessment exercise. All responses are submitted via the provided Google Doc template.

Tools: Google Docs (AI Ethics Impact Assessment template)

Participants complete a two-part summative assessment that integrates skills from across all seven modules. The summative lab serves as the capstone output. It requires participants to build, fine-tune, and deploy an ML or LLM-powered solution end-to-end, with explicit attention to ethical considerations.