Dr Seani Rananga is attending the Deep Learning Indaba 2026 in Lagos, Nigeria. Here is her reflection from Day 4 of the conference.
Day 4 of the Deep Learning Indaba was another inspiring day of learning, sharing research, and celebrating the growing African AI ecosystem.
The morning began with an insightful keynote by Prof. Vukosi Marivate titled “Rethinking AI Benchmarks for African Languages.” He challenged us to rethink what our AI benchmarks actually measure, arguing that evaluation for African languages must go beyond global standards to reflect our linguistic diversity, dialects, cultural context, and oral traditions. It was a timely reminder that meaningful progress in African AI requires evaluation methods that truly represent the communities we aim to serve.
One of the standout sessions today was the Community Session, celebrating the grassroots AI communities that are expanding access to AI education, building local capacity, and creating opportunities for the next generation of researchers and practitioners. Hearing from community leaders including Kadidja Janny Pombot Fall, Rahma Boghale, Lauriane Mbagdje Dorenan, Samkelo Msibi, Dr. Brando Okolo, Faruq Afolabi, Favour Falade, Victor Ogundele, Aanu Oyeniran, Chinonyelum Rosemary Igwe, Petra Agien, Gilles Quentin Hacheme, and moderator Paul Kennedy highlighted the importance of collaboration, sustainability, and investing in local AI ecosystems across Africa.
The conference also reflected on the success of Research in Africa Day, which showcased more than 200 research presentations across IndabaX chapters, spotlight papers, African datasets, and general AI research. It was inspiring to see the depth, diversity, and impact of AI research taking place across the continent.
The highlight of my day was presenting my PhD research poster on multilingual misinformation detection in isiZulu and Sepedi, using COVID-19 as a pilot study. It was a privilege to discuss my work with researchers from across Africa and receive valuable feedback and thought-provoking questions.
My research presents a multilingual misinformation detection framework that explores:
Every discussion today reinforced how much innovation is happening across Africa and how powerful collaboration can be in solving challenges that matter to our communities.
Grateful for another incredible day at the Deep Learning Indaba and looking forward to what the rest of the week has in store.
Follow our journey at Deep Learning Indaba 2026 and beyond:
Moyahabo Rabothata, a PhD student supported by the AI4D African Languages Lab, is attending the Deep Learning Indaba 2026 in Lagos, Nigeria. Here is her reflection from Day 3 of the conference.
Day 3 of Deep Learning Indaba was both thought-provoking and inspiring. The day began with a powerful keynote by Dr. Melanie Mitchell, who challenged us to look beyond the hype surrounding Artificial Intelligence. Rather than focusing only on AI’s promise, she examined the realities of how the field is evolving, highlighting how the current AI landscape is increasingly shaped by a small number of powerful organizations. Her talk encouraged us to question whether today’s rapid advances are truly delivering measurable benefits for society or primarily reinforcing existing power and commercial interests. The session was followed by an engaging fireside chat moderated by Raesetje Sefala from the DAIR Institute, which further explored these important conversations around responsible and inclusive AI.
One of the most inspiring moments of the day was celebrating African research excellence through the Alele-Williams Master’s Award and the Kambule Doctoral Award. Congratulations to Akinbobola Adegboyega (Nigeria) for his remarkable work on evaluating machine translation for low-resource tonal languages, placing emotion and cultural meaning at the centre of translation quality. Equally inspiring was Dr. Everlyn Chimoto (Kenya), whose doctoral research is advancing natural language processing for African languages while making AI more accessible under real-world African resource constraints. These achievements are a reminder that world-class AI research is being driven from within Africa and is addressing challenges that matter most to our communities.
A personal highlight of my day was visiting the Mila booth and had the opportunity to catch up with Prof. David Ifeoluwa Adelani to discuss his work on IrokoBench and AfriMGSM. Since my own research focuses on developing a mathematical reasoning benchmark for low-resource South African languages, it was incredibly motivating to engage with someone whose work features so prominently throughout my literature review. Conversations like these remind me why attending conferences like Deep Learning Indaba is so valuable, they create opportunities to learn directly from researchers whose work is shaping the future of African AI which is embodied in this year’s theme “Ìmọ̀ Wa” (our shared knowledge).
My day concluded with the Masakhane Community Dinner at Four Points by Sheraton. The evening opened with a panel discussion on the state of African NLP, reflecting on the journey so far, the challenges facing the AI ecosystem, and the exciting opportunities ahead for African researchers and innovators. It was a fitting end to a day filled with learning, meaningful conversations, and new connections with people who share the vision of building AI that truly serves African communities.
Day 3 reminded me that impactful AI is not only about building more powerful models, it is about building knowledge, communities, and solutions that reflect our languages, cultures, and realities. Every conversation reinforced my belief that the future of African AI will be shaped by researchers who are bold enough to tackle our unique challenges while collaborating to create technologies that leave no language or community behind.
Follow our journey at Deep Learning Indaba 2026 and beyond:
Nontokozo Manukuza, an MSc student supported by the AI4D African Languages Lab and a research assistant on our Data.org Language Playbooks project, is attending the Deep Learning Indaba 2026 in Lagos, Nigeria. Here are their reflections from the first two days of the conference.
The first two days at the Deep Learning Indaba 2026 in Nigeria have been an incredible experience. It is my first time visiting Nigeria, and I have already learned so much.
Day 1 was mainly about settling in after travelling. The welcoming event gave us a chance to experience Nigerian culture through the food, music, and hospitality. We also reflected on this year’s theme, “Ìmọ̀ Wa” (Shared Knowledge), which set the tone for the entire event. It highlighted the importance of collaboration, learning from one another, and building knowledge that can benefit communities across Africa. One of the sessions on reading and writing research papers stood out because it shared practical techniques for understanding research papers and writing stronger scientific papers — skills that will be very useful in my academic journey.
Day 2 was even more inspiring. Dr. Catherine Nakalembe’s keynote showed how GeoAI can help solve African challenges such as agriculture, disaster management, and environmental monitoring. Her message about building AI solutions that are designed for African contexts reminded me that technology has the greatest impact when it addresses the realities of the communities it serves.
Another highlight was Prof. David Adelani’s session on building African-centric large language models for African languages. He emphasized the importance of high-quality datasets, careful data curation, strong evaluation benchmarks, and developing models that truly understand African languages. His work strongly connects with my own Master’s research on teaching large language models to correctly interpret isiZulu idioms while preserving their cultural meaning.
What made this even more meaningful was how well it connected with the cloud GPU infrastructure tutorial that followed. Prof. Adelani’s session showed us what it takes to build better African LLMs, while the tutorial demonstrated how we can train and scale those models using dedicated computing resources. Building and fine-tuning modern LLMs requires powerful GPU infrastructure, and the tutorial highlighted why moving beyond platforms like Google Colab to dedicated GPU environments is essential for training larger models, experimenting more efficiently, and deploying AI solutions at scale. Together, these sessions reinforced that advancing African LLMs requires both high-quality language resources and the computing power to train them effectively.
Overall, these first two days have motivated me even more. Every session has reminded me of the importance of learning, sharing knowledge, and building AI solutions that address African challenges. I am looking forward to learning even more in the coming days.
Follow our journey at Deep Learning Indaba 2026 and beyond:
The countdown is officially on for the Deep Learning Indaba 2026 at Pan-Atlantic University in Lagos! As the Data Science for Social Impact (DSFSI) research lab at the University of Pretoria, our mission is to advance AI and Data Science research that directly serves our communities, languages, and sociotechnical contexts.
This year, members and researchers from the DSFSI lab will be represented across the main plenary stage, accepted publication tracks, interactive workshops, hands-on tutorials, poster showcases, and peer-review panels under the conference theme Sovereign Intelligence: Africa’s Path in a Frontier AI World.
Here is where you can find and connect with the DSFSI team during Indaba week!
Lab Principal Investigator Prof. Vukosi Marivate will be delivering a main plenary keynote address to the full conference assembly on the critical challenges surrounding AI evaluation in African contexts:
| 🗓️ When: Wednesday, 5 August 2026 | 09:00 (GMT+1) |
We are thrilled to share that a paper authored by Prof. Vukosi Marivate has been accepted into the Deep Learning Indaba 2026 Publication Track:
Postdoctoral Fellow and Research Affiliate Dr Idris Abdulmumin is co-organising two major workshops and leading a hands-on tutorial focused on community-driven language resources and practical data extraction challenges:
Members of our lab are presenting cutting-edge research at the main Indaba Research Showcase:
We are also proud to share that Moyahabo Rabothata (PhD Candidate) will be serving as a Poster Judge at this year’s conference, helping to review, evaluate, and mentor emerging researchers showcasing their work from across the continent.
🤝 Connect with Us! If you are attending #DLI2026 in Lagos, make sure to catch Prof. Marivate’s keynote and paper presentation, drop by Dr Abdulmumin’s workshops and tutorial session, visit Nontokozo and Dr Rananga’s posters during the showcase, and connect with our team.
Follow our updates throughout the week right here on Substack or visit the DSFSI Website to learn more about our ongoing projects!
]]>On June 28th, during the Society 5.0 Conference 2026, DSFSI PhD candidate Miehleketo Mathebula participated in a unique doctoral symposium held against the scenic backdrop of Kruger National Park. The event provided a rare opportunity for deep academic mentoring and high-level technical critique, framed by a dual-perspective feedback structure: each presenter was evaluated by a two-supervisor panel—one specialist from their direct field and one from an entirely different academic background.
Miehleketo presented his doctoral work focused on SA-FINESG, an AI-enabled framework designed to provide ESG (Environmental, Social, and Governance) financial intelligence for South African and emerging markets.
His research addresses a critical bottleneck in the current landscape: ESG information is often fragmented across corporate reports, policy documents, and sustainability disclosures. The goal of his work is to transform this scattered, unstructured data into structured, machine-readable, and policy-aware intelligence using a sophisticated combination of data engineering, NLP, transformer-based models, retrieval mechanisms, and knowledge graph representations.
A defining moment of the symposium was the shift from viewing his work as a “technical pipeline” to framing it as a robust research artifact. While the initial focus was on the system’s ability to collect, clean, and process data, the panel encouraged a deeper theoretical positioning rooted in an adapted Design Science Research (DSR) approach.
The feedback highlighted several critical dimensions for elevating the contribution of SA-FINESG:
This shift ensures that Miehleketo’s thesis stands as both a technical solution and a theoretical contribution to how we handle complex, high-stakes data in the South African context.
Beyond the rigorous academic critique, the setting of the Kruger National Park provided an unparalleled environment for networking. The relaxed atmosphere encouraged open dialogue between PhD candidates, supervisors, and researchers from diverse institutions.
From the unique experience of spotting wildlife during the drive to the vibrant atmosphere of the cocktail networking sessions—including the organized viewing of the Bafana Bafana vs. Canada World Cup match—the symposium succeeded in fostering a community where research is not just developed in isolation, but within a socially relevant and human-centered academic ecosystem.
We look forward to seeing how these insights into SA-FINESG and the Design Science framework shape the next chapters of Miehleketo’s doctoral journey!
Stay connected with our work:
In the world of academic research, we often talk about milestones in terms of papers published, citations gathered, or grant funding secured. But last night, our team at the Data Science for Social Impact (DSFSI) lab experienced a milestone of a completely different scale.
At the annual NSTF-South32 Awards—affectionately known as the “Science Oscars” of South Africa—the TextAugment team walked away with the inaugural NSTF-SADiLaR Research Software Award for Human Language Technologies (HLT).
If you missed the live broadcast, you can watch the exact moment our team received the trophy here on the NSTF YouTube Broadcast at 03:25:46.
This award is a massive honor for our lab, but more importantly, it is a validating moment for the entire Human Language Technology and Natural Language Processing (NLP) community in South Africa and across the African continent.
One of the biggest hurdles in training robust machine learning models for low-resource languages (like isiZulu, Sesotho, or Yoruba) is data scarcity. Traditional AI models require mountains of clean, labeled data to perform well. If that data doesn’t exist, these languages are effectively locked out of the modern AI revolution.
This is where TextAugment comes in.
Developed as an open-source library, TextAugment focuses on generating synthetic training data to help build more robust machine learning models for under-resourced languages. What started as a focused research tool has grown into a vital piece of global language infrastructure:
You can explore the code, contribute, or implement it in your own projects on our GitHub repository.
What made the night exceptionally sweet was looking across at our fellow nominees in this inaugural category. The other finalist was Dr. Herkulaas Combrink and his team from the University of the Free State (UFS), nominated for their pioneering and incredibly vital work on South African Sign Language (SASL).
This was more than just a friendly competition; it was a proud family reunion. Herkulaas was co-supervised by Prof. Vukosi Marivate during his academic journey.
To stand on a national stage side-by-side with a former student and collaborator—both of us championing linguistic equity in different, yet equally critical, areas of language technology—was the ultimate proof of our core philosophy.
Ideas must travel, and they travel best when we invest deeply in the people carrying them.
For a long time, the academic system has struggled to recognise software development as “core” research. Code was often treated as a byproduct of a paper, rather than a primary research output in its own right.
But as the chairperson of the NSTF noted during the gala, software is the invisible engine enabling modern groundbreaking research across all scientific fields—including the humanities and social sciences.
By partnering with the South African Centre for Digital Language Resources (SADiLaR) to launch this new award category, the NSTF has sent a clear message: the software tools we build to study, preserve, and revitalize our languages are critical scientific infrastructure.
An achievement like this is never the work of a single person. It is the result of an ecosystem that began with a spark of inspiration and grew through continuous collaboration.
Our deepest gratitude goes out to:
We also owe an immense debt of gratitude to our families. The late nights, the times away from home, and the cognitive load of chasing these milestones are heavy, and we could not do any of this without your patience and love.
This award is not a finish line; it is a catalyst.
We are incredibly proud of the grassroots African NLP community. Across the continent, researchers are using our tools, creating their own, and actively fighting against digital language extinction. Our collective mission remains clear: to build robust, equitable language infrastructure for all.
As we look to the future, we hope to scale this impact through our newly established platform: the African Institute for Data Science and AI (AfriDSAI) at UP. AfriDSAI is designed to serve as a hub for researchers across UP, South Africa, and the broader continent to build a unified ecosystem dedicated to scientific excellence and real-world societal impact.
To the entire TextAugment team, our collaborators, and the global community of researchers using our library—this award is for you. Let’s keep building.
To read more about all the incredible winners from this year’s awards, check out the official NSTF Current Winners page.
Stay connected with our work:
To keep ourselves honest as a lab, we’ve written down eight rules that guide how we use AI in our research. This is the DSFSI Lab AI Manifesto.
1. I will work with the AI, not for it. The moment you outsource the fundamental reasoning that makes you a researcher, you cease to lead the inquiry. You may use AI to accelerate workflows, automate data cleaning, or suggest code snippets. But keep your human judgment in the loop. If the AI does the thinking, you aren’t a researcher; you are a prompt-operator. TIP: Always be able to explain the “why” behind a result without the help of an LLM.
2. I will prioritise process over product. In an era of instant generation, the “output” (the paper, the code, the summary) is easy. Learning is hard. If you use AI to bypass the struggle of understanding a complex concept, you have traded your long-term intelligence for a short-term deliverable. TIP: Use AI to test your understanding (e.g., “Explain this to me like I’m a peer”), but never to bypass the initial struggle of comprehension.
3. I will think about who is watching. Data is never neutral, and privacy is never free. As we build models, we must remain hyper-aware of the data pipelines and the power dynamics inherent in them. Who owns the data we use? Whose privacy are we compromising for the sake of a benchmark? TIP: Audit your datasets for consent, representation, and the long-term implications of data extraction.
4. I will protect the “slow thinking” space. Deep insights rarely come from rapid-fire prompting. They come from “slow thinking” — the deliberate, messy, unhurried process of synthesis. AI is built for speed; research requires depth, search, and critique of different literatures, not just copy, paste, and paraphrase. TIP: Keep a physical notebook. Sketch architectures, write out logic flows, and draft hypotheses with pen and paper before touching a keyboard. Give your brain the friction it needs to build permanent neural pathways.
5. I will keep building my own training data. Your unique observations, your field notes, and your lived experiences in the community are the only data points that cannot be scraped from a server. Do not let your expertise become a derivative of a pre-trained model. TIP: Prioritise primary research and ethnographic observation. Your “ground truth” should come from the real world, not just a digital proxy.
6. I will value human friction. AI is designed to be frictionless, but scientific progress often requires the friction of human debate, disagreement, and messy collaboration. A chatbot will rarely challenge your fundamental assumptions; a colleague will. TIP: Seek out intellectual conflict. Debate your findings with humans. The most profound breakthroughs happen in the heat of interpersonal discourse.
7. I will practice radical skepticism. In NLP, “hallucination” is a feature, not a bug. Treat every AI-generated insight, citation, or summary as a hypothesis that requires rigorous empirical verification. Never assume a model is “correct” simply because it is confident. Verify and benchmark your results against the existing body of knowledge, and if your results are far off, be able to rigorously explain why. TIP: Verify the “unverifiable.” If an AI cites a paper, find the paper. If it explains a concept, check it against a textbook.
8. I will design for social impact, not just accuracy. A model can be mathematically “accurate” while being socially catastrophic. Our goal is not just to minimise loss functions, but to minimise societal harm. We do not build for the sake of the model; we build for the sake of the people the model affects. TIP: Before deploying or publishing, ask: “Who does this model exclude, and who does it inadvertently empower?”
This manifesto isn’t a set of restrictions on using AI — it’s a commitment to staying researchers, not prompt-operators, as these tools become part of everyday practice. We’ll keep revisiting it as the field moves.
Read, comment, or download the manifesto as a PDF via our Google Doc.
© Data Science for Social Impact Lab, University of Pretoria. Licensed under CC-BY. Inspired by Joanna Stern’s “Five Rules for Living in an AI World.”
]]>As we navigate an era of information overload, the ability to distinguish between truthful knowledge and systemic misinformation is no longer just a technical challenge — it’s a societal necessity. Our session explores how Retrieval-Augmented Generation (RAG) combined with Knowledge Graphs can provide the grounding necessary for reliable AI applications.
The Multi-Lingual Frontier We showcased a prototype built to tackle COVID-19 misinformation, specifically optimised for English, isiZulu, and Sepedi. This highlights our commitment to ensuring that the benefits of LLMs are accessible in local contexts.
Theory Meets Practice Fiskani Banda led a deep dive into the theoretical framework of RAG and Knowledge Graphs, demonstrating how these systems can address knowledge gaps in specialised sectors, such as South African agriculture, where general-purpose models often struggle.
Hands-on Innovation Beyond the theory, participants got their hands dirty in a collaborative session (with partners from CSIR and UP) covering:
This workshop was a true community effort, featuring contributions from the CSIR, University of Pretoria (UP), and University of the Free State (UFS) teams.
We were also joined by other team members including Dr Abebe Tegene, Nontokozo Manukuza, Moyahabo Rabothata, and Hawa Ibrahim, who all contributed to the session.
Stay connected with our work:
Here are the papers from the lab being presented this week:
Dr. Abdulmumin will be talking at the Beyond Alignment: Transdisciplinary Conversations on Human-AI Futures (BATCH) Workshop. His talk, titled “Perspectives on Alignment from Low-Resource Languages,” unpacks the unique hurdles and necessary solutions for aligning AI systems in contexts where linguistic data is scarce.
To continue fostering community-led innovation, Dr. Abdulmumin has also co-organised two shared tasks for SemEval-2026:
Dr. Modupe is on the ground in San Diego representing the AI4D African Languages lab, an initiative hosted at DSFSI dedicated to creating digital innovations that contribute to Africa’s digital transformation. The project focuses on foundational language models, responsible AI, and high-impact NLP applications tailored for African linguistic contexts.
Dr. Modupe is happy to meet those interested in our project, share insights on our ongoing capacity-building efforts, and discuss how we can work together to bridge the digital divide.
Let’s Connect in San Diego! We want to hear from you. If you are attending ACL 2026, please reach out to Dr. Abdulmumin and Dr. Modupe after their sessions or via our team contact page to schedule a meetup.
Stay connected with our work:
RSVP https://forms.gle/Dz1Uhcou8PvhExGf6
Bio Jean Louis Fendji is an Associate Professor of Computer Science at the University of Ngaoundéré, Cameroon, where he leads the Centre for Research, Experimentation and Production (CREP) at the School of Chemical Engineering and Mineral Industries (EGCIM). He holds a PhD (Dr.-Ing.) in Computer Science from the University of Bremen, Germany.
His research leverages artificial intelligence, data justice, and optimization techniques to advance Sustainable Development Goals, with a focus on rural connectivity, precision agriculture, and digital inclusion for low-literacy populations speaking oral African languages. An Iso Lomso Fellow at the Stellenbosch Institute for Advanced Study (2024-2026) and a former fellow at the Hamburg Institute for Advanced Study (2025), Fendji is a Co-Principal Investigator for the EU Horizon project DIGITAfrica. He also contributes to national tech policy as a member of the ICT and Artificial Intelligence Commission within Cameroon’s Ministry of Scientific Research and Innovation.
Abstract
Africa contributes just 2% of global AI training data — yet the dominant institutional response remains the same: build more datasets. This presentation argues that this response, however well-intentioned, mistakes a structural problem for a logistical one. Drawing on a directed content analysis of nine African national and continental AI policy documents published between 2021 and 2025, we demonstrate that African AI strategies systematically address dataset creation while remaining blind to datafication — the continuous, infrastructure-embedded processes through which data is generated as a structural by-product of digital systems. The empirical finding is stark: across 288 coded segments, Dataset Creation Language outnumbers Datafication Language by 2.4:1, and not one of the nine strategies uses the word “datafication” or any direct conceptual equivalent. Seventy-six absence markers document the specific passages where datafication framing is missing despite contextual expectation. We examine what this blind spot looks like in practice — including the Ethiopian exception, where a DFL-dominant strategy plans IoT deployments across six sectors and 600 petabytes of sovereign AI storage, yet never connects these investments to AI training data production. The presentation closes with a datafication-first framework built on three reorienting principles and a tiered infrastructure model designed for Africa’s resource realities.
Stay connected with our work: