Advanced NLP & Large Language Models

Course outline and learning expectations

This is a tutor-led course. The outline shows what your tutor will cover; teaching materials and examinations are provided directly to enrolled students.

What you will learn

• Explain tokenisation, embeddings, self-attention, transformer blocks, context windows and text generation at an application level
• Design and evaluate structured prompts for complex NLP tasks
• Choose appropriately between prompting, retrieval-augmented generation and fine-tuning based on data, cost, privacy and quality requirements
• Integrate a large language model API into a Python application
• Build a basic RAG pipeline that ingests, chunks, embeds, retrieves and cites local documents
• Measure answer quality, groundedness, latency and token cost using a repeatable test set
• Apply safeguards for hallucination, prompt injection, privacy and responsible AI use
• Design an LLM deployment suitable for African bandwidth, connectivity and budget constraints
• Deliver a capstone study assistant using authorised public curriculum materials, UNEB sample or past papers and learner notes

Modules

1. 1

   How Large Language Models Work

   Build an application-level understanding of transformer-based language models.

   NLP foundations and the move to large language modelsTokenisation and embeddingsSelf-attention and transformer blocksPretraining and instruction tuningContext windows and text generationModel capabilities and limitations
2. 2

   Advanced Prompt Engineering and Evaluation

   Design prompts systematically and test whether they meet defined success criteria.

   Instruction hierarchy and prompt structureZero-shot and few-shot promptingContext and examplesStructured outputsPrompt templates and versioningTest sets and evaluation rubrics
3. 3

   Prompting, RAG and Fine-Tuning Decisions

   Select the most suitable adaptation method for a task without unnecessary cost or complexity.

   When prompting is sufficientWhen retrieval is requiredFine-tuning concepts and suitable use casesTraining-data qualityPrivacy and intellectual-property considerationsCost, maintenance and performance trade-offs
4. 4

   Building Applications with LLM APIs

   Develop a dependable Python application that communicates with a commercial LLM API.

   API authentication and secret managementRequests, responses and conversation stateStreaming and structured responsesRate limits, retries and error handlingLogging and basic observabilityProvider abstraction and application prototyping
5. 5

   Retrieval-Augmented Generation with Local Documents

   Ground model answers in relevant documents that can be updated without retraining the model.

   Document ingestion and cleaningChunking strategiesEmbeddings and vector searchRetrieval and prompt assemblySource citations and answer groundingRAG evaluation and common failure modes
6. 6

   Cost, Latency and Connectivity for African Deployment

   Plan LLM services around practical infrastructure and affordability constraints in Uganda and Africa.

   Token and request cost estimationModel selection and routingCaching and batchingLatency measurement and streamingLow-bandwidth user experienceIntermittent-connectivity fallbacksMonitoring usage without hard-coding provider prices
7. 7

   Safety, Hallucination and Reliability

   Reduce avoidable errors and protect users, data and source documents.

   Hallucination and groundedness checksPrompt injection and unsafe inputsAccess control and data minimisationUganda privacy and data-protection considerationsHuman review for high-impact outputsRed-team testingReliability evaluation
8. 8

   Capstone - UNEB-Aligned Study Assistant

   Build and demonstrate an LLM-powered study assistant grounded in authorised learning materials.

   Define learners and supported subjectsUse NCDC curriculum materials and authorised UNEB sample or past papersPrepare and index notes and documentsImplement the RAG question-answering workflowReturn source-aware answersTest accuracy, latency and costDocument limitations, safeguards and deployment choices