I build the parts that have to hold a latency budget: CUDA and WebGPU kernels, process-level audio capture, voice conversion, avatar rendering, and the WebRTC path that puts all of it on screen. Then I build the product on top.
Engineering by m96-chan — Yusuke Harada, Representative Director of DEVenus Ltd. (Shinjuku, Tokyo). Corporate contracts and invoicing go through the company; smaller one-off consultations are welcome directly.
One character illustration in. A live, speaking, breathing presenter out — rendered locally, no cloud, straight into OBS.
Most "AI VTuber" demos are a cloud API with a face on it. This one runs on the machine in front of you — because I own every layer, from the model port to the virtual microphone to the stream transport.
Turns a single front-facing illustration into an animated presenter that reads a PDF deck aloud, page by page. Talking Head Anime 4 ported to Rust + candle (CUDA / Metal / CPU from one codebase), with a per-character distilled student model for the real-time path.
Real-time zero-shot voice conversion in pure Rust. A few seconds of reference audio, and a virtual microphone delivers the character's voice to Discord, Zoom or OBS. Five engines behind one interface, no Python and no CUDA Toolkit at runtime.
A C++ OBS Studio plugin giving universal WebRTC input and output — low-latency two-way video and audio in and out of the scene graph. Paired with ProcTap when a single application's audio needs to be captured without touching the rest of the desktop.
Status, honestly: the browser app animates, speaks and reads a deck today, with three characters distilled; slide transitions and mic-driven lip sync are still being built, in the open. If you want this shape of system for a product — a presenter, a receptionist, a companion, a support agent with a face — the hard parts are already solved and the integration is the work.
Four things I do repeatedly, each with shipped code behind it. If your problem sits between two of them, that is usually exactly where I am most useful.
End-to-end: character animation, voice, lip sync, layout, and the path onto a stream or into your app. Local-first when privacy or cost rules out a cloud round trip.
Proven in VPresentation · babiniku.rs · voxshot
When the model is fast enough on paper and too slow in production. Kernel-level work on the architecture you actually ship on, from datacenter cards down to a browser tab or a phone.
Sub-100 ms systems where capture, inference and playback all have to fit in the same budget — and where the interesting failures are in the OS, not the model.
Proven in ProcTap · PyETWkit · MeetScribe
Architecture review, technology selection, and helping a team actually adopt agent-assisted development instead of just installing it. Fifteen years of shipping, including CTO and technical-lead roles.
Background CTO at a product company · technical lead across a 4-vendor platform consolidation · offshore vendor control (India / Vietnam)
Pick whichever is closest. Scope, timeline and price come out of a first conversation — there is no form to fill in and no minimum engagement to qualify for.
One hard question, one call. "Can this run in 40 ms?" "Is this the right GPU?" "Why does this stutter only in the browser?"
A short, bounded build that answers whether the thing is possible on your hardware and your budget — with numbers, not opinions.
Building the system, or the difficult layer of it, alongside your team. Delivered as code you own, with the reasoning written down.
An ongoing retainer: architecture review, technology calls, and raising the team's floor over months rather than sprints.
You do not need a spec. Four lines is enough for me to tell you whether I am the right person — and if I am not, I will say so and point you somewhere better.
Everything here is a library or tool I wrote and maintain. Several ship to package registries; several have live demos you can run in a browser right now.
Runs Microsoft's BitNet b1.58 ternary LLMs on WebGPU — custom WGSL kernels, no WASM and no server. Ships to npm, crates.io, PyPI and Maven Central with TypeScript, Rust, Python, Swift, Java and C bindings.
Capture audio from one specific process by PID, built on WASAPI process loopback. The library that makes "record only the game, not the call" a two-line Python program. Published on PyPI.
Hourly-updated tracking of AI/ML papers, models, repositories and trends, pulled from arXiv, Hugging Face and GitHub. My own standing observation post, made public.
A minimal GPU runtime for Python: hand-written CUDA kernels, explicit memory management and LLM inference without a heavyweight framework underneath. Ampere through Blackwell, including SM120 on the RTX 5090.
A WGSL → NPU transpiler: compiles WebGPU shading language into native NPU formats, so one shader targets the accelerator that is actually in the device. Backends for Qualcomm, MediaTek, Samsung, Rockchip, ARM Ethos, CEVA, Intel, AMD, CoreML, ONNX, TFLite and StableHLO.
A high-performance Event Tracing for Windows consumer for Python, with a Rust core bound through PyO3 — real ETW throughput without dropping into C++ yourself.
Browser-first zero-shot text-to-speech and voice cloning for JavaScript — WebGPU plus ONNX Runtime Web. No Python, no backend, no API keys. This is the voice inside VPresentation.
KDE-Plasma-style window tiling for macOS: a visual tile editor, Shift+Drop window placement and gap-drag resizing — with no SIP disabling required.
Turns a spare Android phone into a GitHub Actions self-hosted runner — ARM64 builds and real-device NPU tests, without root and without Termux.
I started on Windows client software in C++ and Win32 — payroll software, installers, the kind of codebase where a data-corruption support ticket lands on the person who wrote the library. Then game servers at CAPCOM, where a live title teaches you what latency and load actually mean, and where I ended up leading server operations.
After that, CTO and technical-lead roles: consolidating four vendors onto one platform, building engineering organizations, running offshore teams in India and Vietnam. Since 2023 I have run DEVenus as its representative director, focused on AI and 3D work.
The thread through all of it is vertical range. I can write the CUDA kernel and the React Native front end that displays its output, and I can explain the trade-off between them to someone who does not write code. Right now that mostly means real-time AI voice and video — currently building GPU inference on SM120 and SM86 kernels for a real-time AI voice chat product, and porting avatar models to Rust in the evenings.
I work in Japanese and English, and I have no allergy to talking to executives.
A rough idea is enough to start. I reply to everything, usually within a day or two — including "this is probably not for you, but…".
[email protected]Japanese or English, both fine. Corporate contracts and invoicing are handled through 有限会社DEVenus; NDAs are no problem.