Muhammad Karage

I built and shipped Twiq, a live Hausa, Yoruba, and Igbo transcription and subtitling product, with a Nigerian radio station onboarded as a customer. At Memperor Tech, I fine-tuned Whisper for Hausa across 250k+ real audio files, producing correct Hausa diacritics and no English hallucinations on test audio. I built distributed JAX/Flax training pipelines with streaming from Backblaze B2 and resumable Orbax checkpointing. I also designed and trained a 9.25M-parameter Hausa ASR model from scratch in PyTorch using Conv2D, Transformer, and CTC. To support this work, I scraped and published a roughly 100k-item Hausa text dataset on HuggingFace. I'm comfortable owning production AI systems end to end: multilingual data infrastructure, model training, FastAPI backends, and real users. I'm looking to contribute these skills on an experienced remote team building production AI systems.