Projects

I co-founded ByteFuse and lead its overall AI and scientific strategy: setting research direction, defining experimental methodology, and establishing quality standards across every ML workstream. Day to day this means mentoring researchers and engineers, architecting multilingual speech models, and owning ML from research prototype to production deployment with real-time multi-model orchestration and mobile-first constraints.

A partnership with Maskew Miller Learning to bring personalised, CAPS-aligned tutoring to South African learners through the channel they already use every day: WhatsApp. Featured on CapeTalk; now used by learners across the country.

The system orchestrates multiple models in real time under mobile-first constraints, grounding responses in curriculum content while keeping latency and cost workable at scale.

Open-source reference implementation accompanying the Interspeech 2023 paper. Includes the consolidation gradient step, template buffer, and evaluation harness for Google Speech Commands and FACC.

Open-source companion to the ICML 2022 Pre-training Workshop paper. Computes RMQM by tracking displacement of embeddings under white-noise injection and PGD adversarial perturbations, giving a structural view of representation quality that goes beyond linear-probe accuracy.