Accepted Tutorials & Skill Sessions

Abstract : As artificial intelligence rapidly advances into the frontier era, questions of sovereignty, fairness, and accountability have become central to Africa’s participation in the global AI ecosystem. Many responsible AI methods were developed in data-rich environments and often rely on demographic labels or assumptions that do not transfer well to African contexts, where data may be sparse, heterogeneous, or culturally nuanced. This tutorial introduces principled and practical approaches for building responsible and sovereign AI systems tailored to African realities. We begin by revisiting core foundations of responsible AI including fairness, transparency, robustness, and accountability, through the lens of low-resource and cross-cultural deployment settings. Participants will learn how bias can emerge across the machine learning pipeline and how to diagnose it using widely adopted fairness metrics such as demographic parity and equalized odds. The tutorial will also highlight emerging directions, including fairness-aware learning without demographic labels and evaluation under distribution shift. Through guided hands-on exercises and case studies spanning health risk prediction, African language technologies, and social media analysis, attendees will gain practical skills for auditing and improving model behavior. By the end of the session, participants will be equipped to design AI systems that are not only accurate but also equitable, trustworthy, and locally grounded. This tutorial directly supports the Indaba 2026 theme by strengthening technical capacity for sovereign, responsible AI development across Africa.

Led by : Blessing Ogbuokiri 

Abstract : As AI tools become integrated into the life sciences, the essential skill for the modern researcher is shifting from running static models to orchestrating "Agentic Workflows." This tutorial moves beyond basic prompting to demonstrate how researchers can use AI agents as collaborative "Lab Partners" to handle iterative, data-driven tasks that call specialised machine learning models. The session provides a hands-on Google Colab experience centred on a critical real-world problem: optimising protein thermal stability. Many life-saving biologicals, such as vaccines and insulin, are fragile and require a strict "cold chain" (2°C to 8°C) that is difficult to maintain in rural or high-temperature regions. By using agents to redesign these molecules for heat resistance, we can improve global health outcomes and industrial efficiency.

Participants will learn to: 

1. Use agents to identify and prioritise biological research targets.
2. Automate data orchestration from public repositories like UniProt.
3. Integrate specialised scientific libraries for structural analysis and stability prediction.
4. Design iterative feedback loops where agents suggest and refine sequences based on calculated metrics.

Participants will leave with a practical framework for using AI agents to navigate complex biological datasets and automate the "drudgery" of research. No prior biology background is required; basic Python proficiency is sufficient.

Led by : Natasha Latysheva ; Sebastian Bodenstein ; Kyle Taylor ; Adam Kosiorek

Abstract : As artificial intelligence systems grow more capable, ensuring they behave in ways consistent with human values and intentions—known as model alignment—has become a central concern in AI safety. This tutorial, presented at the Deep Learning Indaba, offers an accessible, hands-on introduction to the core concepts and practical techniques for aligning modern AI models. We begin by motivating the alignment problem through real-world examples of misaligned behaviour in deployed systems, illustrating how models can produce harmful, biased, or unintended outputs even when achieving strong benchmark performance. We then introduce foundational alignment concepts, including the distinction between outer and inner alignment, reward misspecification, and goal misgeneralization. The tutorial surveys the most widely adopted alignment techniques—reinforcement learning from human feedback (RLHF), direct preference optimisation (DPO), and constitutional AI—explaining each method's intuition, workflow, and trade-offs in a manner accessible to participants with a working knowledge of deep learning. A significant portion of the session is dedicated to practical steps: we walk-through how-to curate preference datasets, fine-tune a language model using open-source alignment toolkits, design red-teaming exercises, and evaluate alignment using both automated and human evaluation protocols. We also address considerations especially relevant to the African AI community, including aligning models to diverse cultural values, handling low-resource languages, and building locally relevant evaluation benchmarks. The tutorial concludes by highlighting open challenges, such as scalable oversight and superalignment, and encourages participants to contribute to this rapidly evolving and globally important field.

Led by :  Michael Mollel ; Winnie Mangeni  ; Innocent Charles 

Abstract: While Google Colab has democratized initial access to machine learning across Africa, its infrastructure limitations, such as session timeouts, lack of persistent storage, restricted environment control, and limited compute, stifle the transition from academic experimentation to production-ready AI. True “Sovereign Intelligence” requires African researchers to graduate from managed notebooks to dedicated, localized cloud infrastructure. This two-hour tutorial introduces participants to the
fundamentals of developing on sovereign African GPU Virtual Machines (VMs). We begin by identifying the “Colab Ceiling” and detailing the core architectural advantages of dedicated VMs, including persistent environments, massive data pipelines, custom CUDA configurations, and specialized workloads like fine-tuning open-source LLMs, Retrieval-Augmented Generation (RAG), and multimodal generative AI media pipelines (image, music, and film). To ensure an interactive experience, every participant will receive a $5 credit to log into an active aghcloud.ai African-hosted GPU instance live during the session. Using the AGH GPU Optimizer tool, attendees will learn how to analyze their specific workload requirements and match them to the most cost-effective hardware. By the end of this tutorial, students and researchers will possess the practical DevOps skills necessary to optimize, deploy, monitor, and scale sovereign AI applications entirely within Africa.

Led by : Chisom Chinwuba  ; Sheriff Issaka  ; Opemipo Durodola 

Abstract: This tutorial offers a deep dive into the mathematics of neural networks, the building blocks of contemporary AI. Starting from an intuitive understanding of 'what it means for a machine to learn,' the tutorial introduces essential concepts such as cost functions and gradient descent, before presenting the artificial neuron and the feedforward mechanism. The centrepiece is backpropagation: participants work through the gradients themselves, guided by slides that reveal the derivations progressively, ensuring active engagement rather than passive reception. The concepts of memoization and tensor dimensions are also addressed. By the end, participants will have a good understanding of how neural networks are trained. The tutorial closes with a live demo of TensorFlow Playground in which we  explore how architecture choices (depth, width, activation functions) shape the resulting decision boundaries. A follow-up skills session, "Code a Neural Network From Scratch" is  a hands-on session that complements the theory.

Led by : Pelumi Victor  ; Catherine Essuman 

Abstract: Modern Transformer models are increasingly powerful, but their real-world usefulness depends heavily on how efficiently they can be trained and served on GPU hardware. This tutorial introduces participants to GPU kernel programming with Triton through the practical goal of building faster Transformer components from the ground up. We will begin with the performance bottlenecks that commonly appear in Transformer workloads, including memory bandwidth limits, attention computation, operator overhead, and inefficient data movement. Participants will then learn the core ideas behind GPU execution, tensor layouts, parallelism, tiling, and memory hierarchy before implementing simple Triton kernels.
The hands-on section will focus on writing, benchmarking, and optimizing kernels for common Transformer operations, gradually building toward attention-related kernels used in modern LLM systems. By the end of the tutorial, participants will understand how custom GPU kernels can improve Transformer performance, how to reason about throughput and memory efficiency, and how to begin applying Triton to their own machine learning systems research and engineering projects.

Led by : Ayomide Odumakinde  ; Katherine Oluwadarasimi Olowookere ; Adeniyi Mayokun 

Abstract: Many African systems are inherently relational. Mobile money transaction networks in Kenya and Uganda, electricity transmission infrastructure in South Africa and Zambia, transportation networks in Lagos and Accra, and agricultural supply chains in Morocco and Egypt all form networks of connected entities, and traditional machine learning methods, which treat samples as independent, often miss the signal carried by these connections. Graph Neural Networks (GNNs) are designed to
learn directly from such structure, propagating information across nodes to capture relational dependencies.
This tutorial introduces the foundations of GNNs for an audience new to graph-based learning. It covers core concepts such as message passing and two foundational architectures, the Graph Convolutional Network (GCN) and the Graph Attention
Network (GAT). A live demonstration walks through building and training a GNN for account-level fraud detection on a subsampled mobile money transaction graph, using the MoMTSim dataset developed at Makerere University. Slides, notebooks, exercises, and reading lists will be released on a public GitHub repository after the session. Participants will leave able to represent graph-structured data, build and train basic GNN models, and recognise where such methods can be applied to challenges across finance, infrastructure, agriculture, and urban systems in their own communities and beyond.

Led by : NATARAM ODUMO ; Nwandu Christian ; Godbless James

Abstract: Getting into ML research with little or no institutional backing is hard. Getting published at top venues without a lab, an advisor, or a GPU cluster feels nearly impossible — and for most people trying from Africa, it has been exactly that. But even under those constraints, it doesn't have to stay that way. This session is built around a simple premise: the path exists, it's navigable, and it can be made legible. We know because we have watched people like us walk it — from self-taught practitioners with no institutional affiliation, to first authors at A* venues, to PhD students at world-class programs and research scientists/engineers at big tech companies. The trajectories are real. The moves that made them possible are identifiable. This session reverse-engineers those moves. We will cover the mindset shifts that matter most when working without institutional support; how to find and approach mentors effectively; how to scope research problems that are tractable with limited compute and no proprietary data; how to read and engage with literature efficiently; and how to write and position work for top venues rather than defaulting to workshops and obscure journals; which matters directly when applying to competitive programs and grants. We will be honest about the structural barriers. They are real, they are unequally distributed, and they compound. But the focus here is on what is actually within your control. The goal is not to inspire. It's to leave you with a clearer map than we had.

Led by : Steven Kolawole 

Abstract: This one-hour skill session shows how capable language models can run locally on phones and laptops through careful model selection, conversion and quantisation. It will start with a short demo of a prebuilt mobile app so everyone sees the end result from the start. We then use the Gemma 3n family or a lightweight alternative (Qwen3.5 0.8B) as a primary case study to highlight size, latency, and capability. The session explains the core idea of quantisation: how lowering numeric precision shrinks model size and memory use while preserving much of the model’s usefulness. I walk through the practical conversion and quantisation steps you need to generate a mobile-friendly GGUF or TFLite artifact, plus common pitfalls and tuning tips. Finally, I show how the quantised model is integrated into a React Native Expo project (prebuilt APKs and code provided), describe distribution options, and point to further resources. Everyone leaves with a checklist, links to the exact scripts and builds used in the demo, and a clear path to experiment with on-device models in their own projects.

Led by : Jesse-Paul Osemeke ; Daniel Onyesoh 

Abstract:To achieve "Sovereign Intelligence," African researchers must not only build innovative models but also master the art of global scientific communication. Many young scholars in Africa possess groundbreaking ideas but face barriers in effectively deconstructing dense AI literature or structuring their findings for top-tier venues. This Skill Session is designed to bridge this gap. We will move beyond theory to provide a systematic "how-to" for both consuming and producing research. The session will cover the "3-pass" reading strategy to extract core contributions from complex papers quickly. On the writing side, it will demystify the scientific storytelling process—from crafting a compelling Abstract to presenting results that meet international standards. By equipping participants with these reusable competencies, we empower African AI agency, ensuring that our local innovations are documented, recognized, and integrated into the global frontier of AI. This session targets undergraduate and early-graduate students, providing them with a concrete blueprint for scientific success.

Led by : Castro Gbêmêmali 

Abstract: Modern AI systems are increasingly deployed in high-stakes environments, yet they remain vulnerable to subtle but powerful adversarial manipulations. From evasion attacks and data poisoning to prompt injection in large language models, these vulnerabilities expose critical weaknesses in today’s machine learning pipelines. This practical session introduces participants to the core concepts of adversarial machine learning and AI security. We will explore how and why models fail, demonstrate simple attack scenarios, and discuss practical mitigation strategies that can be applied in real-world systems. The session emphasizes building a security mindset for AI development, encouraging participants to think beyond accuracy and consider robustness, reliability, and long-term system resilience. Through guided examples and discussion, attendees will gain foundational tools to evaluate and strengthen AI systems in resource-constrained and emerging technological environments.

Led by : Manar Hamed