Data Augmentation and Evaluation Quality in Low-Resource Kikuyu Speech Recognition: A Whisper Case Study
Whisper Kikuyu ASR: 25.01% WER on a quality-screened test subset, plus augmentation and evaluation-quality findings.
Read article →Writing
The long overdue solarification of Africa
Africa gets more sun than anywhere on earth, pays 5x more for electricity than it needs to
Read article →AI Agents Need Their Own Infrastructure
An AI agent bypassed its tools and called the GitHub API directly.
Read article →Projects
Kikuyu Speech Recognition
Fine-tuned Whisper large-v3 on approximately 465.9 hours of Kikuyu speech. Audited transcription quality: 34.09% WER on the original test set and 25.01% on a retrospectively screened subset.
View project →Soma Agent
Agentic document search using a ReAct loop that reasons about what to read instead of relying on vector similarity. Small embedding model (90MB), agent reasoning compensates. 50% latency reduction over naive RAG.
View project →PromptGuard
Fine-tuned DeBERTa on ~51k examples for prompt injection detection. Deployed to Hugging Face in 16-bit and 8-bit variants. 3.33ms CPU / 0.44ms GPU inference.
View project →