Vali Neagu publishes open-source software as fspecii, including music-generation interfaces and OpenFelix, a voice-first Mac agent. He also runs web studios and experiments with local inference and Apple Silicon, connecting model capabilities to applications people can operate.[2][1][3]
Start with Ace Step Ui →
X Career
Neagu’s portfolio brings these experiments together with web-development work. He runs Web Design Studio London and the Romanian agency websitefirma.ro, while publishing local AI tools under his GitHub account. The distinction between an interface and the model beneath it is central to this collection: a useful studio needs controls, history, audio selection and deployment work as well as a model capable of generating sound.[1][2]
Contributions
Local tools for music and agent interaction
ACE-Step UI provides a studio interface around the ACE-Step music model, with song parameters, reference audio, queues and a library of generated work. HeartMuLa Studio wraps HeartLib with a similar workflow, including selecting a short reference-audio region and adjusting its influence. These applications make existing model systems easier to operate. OpenFelix applies that product focus to a Mac assistant, combining voice and text input, screen context, skills and scheduled tasks with local or configured cloud models.[5][6][3]
Exploring a less accessible compute path
ANE Trainer investigates training through Apple’s Neural Engine, an accelerator normally exposed for inference rather than general training. The repository describes using reverse-engineered private interfaces and compiled compute graphs for forward and backward passes. Its weight updates use shared surface inputs so kernels do not need recompilation after every change. This is an experimental implementation for studying hardware behavior, with requirements and measurements tied to its documented Apple Silicon setup.[4]
Tools, writing & talks
Tools
Start here
Ace Step Ui
github.comA studio interface for ACE-Step with generation controls, reference audio and history.
Start with the feature tour to distinguish the studio interface from the model underneath. Running it requires the documented local model and compute setup; check those requirements before downloading weights.
Interface developer; built on ACE-Step[5]Original source- ANE Trainergithub.com
An experimental transformer-training implementation using reverse-engineered Apple Neural Engine interfaces.
Repository owner[4]Original source - OpenFelixopenfelix.com
A voice-first macOS assistant combining local or cloud models, screen context, skills and scheduling.
Creator[3][2]Original source - HeartMuLa Studiogithub.com
A HeartLib-based music interface with waveform selection, reference-audio influence and generation history.
Interface developer; built on HeartLib[6]Original source