Focus Track
AI & Machine Learning
Building practical, ethical AI for campus life in South Africa: multilingual, privacy-aware, and grounded in student outcomes.
AI vs ML vs Deep Learning
- •AI: systems that mimic reasoning (chat assistants, routing decisions).
- •ML: patterns learned from data (mark predictions, study recommender).
- •Deep Learning: neural nets handling complex signals (speech, images).
- •Student example: timetable assistant that auto-builds a schedule from module data.
Applied on Campus
- •Study recommender: suggest past papers/notes based on module outcomes.
- •Smart reminders: track deadlines and send multilingual nudges.
- •Labs and queues: forecast lab congestion from historical usage.
- •Accessibility: captioning and translation for recorded lectures.
Languages & Data in SA
- •Prioritise English, Afrikaans, isiXhosa for Stellenbosch cohorts.
- •Curate clean, representative datasets; avoid scraping without consent.
- •Balance dialects; avoid overfitting to urban-only data.
- •Document data sources, gaps, and collection ethics.
Ethics & Safety
- •Bias: audit outputs for demographic skew; run red-team prompts.
- •Privacy: no personal marks or IDs in prompts; minimise PII retention.
- •Hallucinations: require citations for generated answers; keep human review.
- •Monitoring: log model inputs/outputs with access controls and rotation.
Model Patterns
- •Classification: email routing (admin vs finance vs academic).
- •Forecasting: venue demand, study room utilisation.
- •Retrieval + generation: campus handbook Q&A with citations.
- •Computer vision: lab equipment usage detection with privacy overlays.
Tools & Skills
- •Python, PyTorch/TF, scikit-learn; notebooks for fast iteration.
- •Vector stores, embeddings, simple RAG for campus knowledge bases.
- •Pipelines: data versioning, eval suites, prompt tests, observability.
- •Deployment: serverless endpoints, GPU/CPU mix, caching outputs.
Ethics & Safety Checklist
Have a rollback path when a model misbehaves.
Store raw data separately from prompts; encrypt at rest.
Explain model limits clearly to users; set expectations.
Ship small, measured improvements; avoid overpromising.