I build and run local AI pipelines for media production — video generation, voice cloning, music generation — and maintain a fleet of self-hosted services. Currently exploring large-model inference on big-memory GPUs (AMD Instinct MI300X / ROCm).
A self-hosted iOS push notification client, forked from Finb/Bark (MIT) and adapted for my own automation stack: home-server alerts, missed-call notifications and machine pings — no third-party push dependency.
LTX-2.3 / Wan 2.2 image-to-video workflows on Apple Silicon (MLX), F5-TTS voice cloning for narration, automated QC and subtitling. Powers two running video channels.
YuE2 open-source music model running on free-tier cloud GPUs (Kaggle T4 x2), producing songs with editable ABC notation. Reproducible notebooks, zero-cost setup.
Benchmark plan: single-card 70B+ LLM inference and 128K-context stress tests on AMD Instinct MI300X (192GB HBM3) with vLLM on ROCm, compared against CUDA baselines. Results will be published here.