
OtwoX African RLHF Foundry addresses a critical gap: Africa has 2,000+ languages, yet none appear in major AI training datasets. Millions of Sudanese, Ugandan, and Chadian speakers cannot contribute to modern AI systems. Built on AMD Instinct MI300X (192GB HBM3) during this hackathon, OtwoX Foundry is a production-ready annotation workspace designed to compete with Scale AI and Outlier.ai — but built specifically for African linguistic diversity. Key Features: - 10+ African languages: Sudanese Arabic, Fur, Zaghawa, Nubian, Dinka, Luganda, Acholi - Voice Annotation: Native speakers record corrections via Whisper ASR on ROCm - Bilingual Interface: Arabic/English with RTL/LTR switching - AI Generation: Qwen2.5-7B via vLLM on ROCm 7.2 - RLHF Dataset Builder: Structured JSONL with quality scores, categories, audio paths - Live GPU Telemetry: Real-time MI300X stats in the UI Technical Stack: - Hardware: AMD Instinct MI300X (192GB VRAM) - Runtime: ROCm 7.2 + Docker - Inference: vLLM 0.17.1 - Model: Qwen/Qwen2.5-7B-Instruct - Speech: OpenAI Whisper (Arabic) - Frontend: Gradio 6.14 Strategic Vision: OtwoX is building data infrastructure for African AI. We are the primary data supplier for companies needing African language training data. This project is submitted for the LINGUA Africa grant (Masakhane + Microsoft AI for Good + Gates Foundation, deadline June 15, 2026) — funding up to $250,000 cash and $400,000 compute credits. OtwoX Foundry: Sovereign African-Language AI, powered by AMD.
10 May 2026

This project was born from direct, costly experience. My 2023 fintech startup processed over £2 million in remittances in a month, and our primary bottleneck for scaling wasn't our technology; it was the 1-3 day settlement delays and partner dependencies. Global B2B payments are a £17 trillion market defined by profound inefficiency. The £17 trillion 'Hurry Up & Wait' problem isn't just a statistic for us; it's a multi-million-pound barrier we've personally hit. Our solution, "Agent-Pay," is the tool I wish I'd had. It's an AI-powered agent built for the 'Payments for Real-World Assets' track. We empower an enterprise to use simple human language, like a voice command, to set complex, conditional payment rules. For example: "Pay our Japanese supplier the moment their goods are confirmed at port." The AI agent handles the entire execution. This is powered by our PaymentRouter.sol smart contract, which functions as a “conditional vault” deployed on the Arc Network. First, the payment (USDC) is transferred to this secure vault, where the funds are held in escrow. The contract is designed to interact only with Arc’s official USDC token, ensuring funds move only per the contract's conditions. The AI Agent operates off-chain, monitoring real-world data, such as a shipping confirmation. Once the condition is met, the agent triggers the backend to call the executePayment function. This is where the power of Arc is fully realised. The instant executePayment is called, Arc’s instant finality means the USDC is released from escrow to the supplier immediately. We haven't just sped up a payment. We have built the mechanism that turns a 5-day, high-risk, manual process into a 3-second, autonomous, and secure settlement.
8 Nov 2025