Open to engineering roles Rust · ML inference · tooling

Research code in. Native software out.

I take models that only run in Python and port them to Rust and Candle, then build the software around them. I did the first Rust port of Kyutai’s Pocket TTS. Right now I’m building Director, a native video engine that agents drive over MCP.

Focus
Model ports, inference
Stack
Rust / TypeScript / Python
Ships as
CLI / Tauri / web / WASM
director / showreel 1920 × 1080 / 30 fps

Every frame painted by Director from one HTML file, with no browser. Score from Magenta RealTime.

Toolchain Rust Candle PyTorch ONNX Runtime WASM Skia Vulkan Tauri Next.js TypeScript

01 / Selected work

Things I built from the first commit.

A video engine, model ports to Rust, and the apps and sites built on them. Most of my work is turning research code into something you can install and run.

pocket-tts
$ pocket-tts serve --int8
✓ model    pocket-tts  candle  cpu
✓ api      /v1/audio/speech  (OpenAI-compatible)
→ "Research code in. Native software out."
✓ streaming  first chunk  RTF 0.33
no Python. one binary. also runs in WASM.
128 stars

02 Model port

Pocket TTS in Rust

The first Rust port of Kyutai’s Pocket TTS.

I rewrote Kyutai’s Python text-to-speech model in Rust on Candle. It ships as one binary with streaming output, int8 weights, an OpenAI-compatible HTTP API, a React web UI, Python bindings and a WASM build for the browser. About 3× real time on CPU and 6× on Metal.

  • Rust
  • Candle
  • WASM

03 More ports

PyTorch to Rust, repeatedly

Speech, TTS and tooling for porting itself.

  • parakeet-candle

    NVIDIA Parakeet TDT 0.6B speech-to-text on Candle, v2 English and v3 multilingual, with token, word and sentence timestamps.

  • pycandle

    A transpiler from PyTorch (via torch.fx) to Candle Rust. It solves einsum, checks shapes and switches to flash attention.

  • ort-chatterbox

    Chatterbox Turbo TTS in Rust on ONNX Runtime: zero-shot voice cloning, FP32/FP16/Q8, streaming.

  • MiniMax H3

    A Candle and PyTorch port of MiniMax H3. Private for now.

ThumbForge's storyboard view: four planned thumbnail layouts (big number, versus, face plus words, before and after) for a brief about a cheap knife versus a chef knife. App screenshot

04 AI product

ThumbForge

A local-first thumbnail studio for Qwen-Image.

Runs Qwen-Image-2.1 on your own ComfyUI. A brief becomes a storyboard of layouts, then drafts, then a feed check that scores contrast, focal point and phone-size legibility before an export that stays under YouTube’s 2 MB limit.

  • Python
  • FastAPI
  • Next.js
  • ComfyUI
The NoFaceTube.com homepage, showing the headline Faceless video that holds up and a preview of the Clean Take app. Live site

05 Web product

NoFaceTube

A reference site for faceless video, built to be cited.

An editorial site with a public quality rubric, a dated tracker of platform rules, and niche guides. Astro content collections, React islands, and a Cloudflare Worker for forms and access.

  • Astro 5
  • TypeScript
  • Cloudflare D1
The ClipFactory demo homepage, showing a vertical clip preview, a clip record panel, and the local-first feature summary. Live demo

06 Video tool

ClipFactory

Long footage in. Vertical clips out. Locally.

Transcript-first selection, multi-beat FFmpeg rendering, and deterministic QA gates. The public demo shows the selection flow in the browser, with no upload and no backend.

  • TypeScript
  • FFmpeg
  • Local-first
The StepUpAI homepage, showing the headline Launch your own slot game and a grid of orange tiles. Live site

07 Studio

Step Up Gaming

The name on my GitHub, and the studio I run as software.

StepUpAI runs the game studio, marketplace and publishing label behind Step Up Gaming. I built the site, the content pipeline and the internal tools that keep it running without a team.

  • Next.js
  • React
  • TypeScript
  • Cloudflare

Also built

  • VNCCS-Desktop A Tauri and Next.js desktop app around the VNCCS ComfyUI pack.
  • magenta-windows A Windows player for Google’s Magenta RealTime music model.
  • comfy-workflows Typed, code-first ComfyUI workflows.
  • tauri-rustdocs-candle Semantic search over rustdoc with a Candle embedding model, plus an MCP server.
  • grits A Git-native issue tracker for AI agents.

02 / About

Most of what I know came from finishing things.

No CS degree. A lot of repositories, and a habit of shipping.

I’m a self-taught engineer working under the name Step Up Gaming, though most of what I build now has nothing to do with games. It started with tools for my own work and grew into model ports, a rendering engine and the products on top of them.

I like the part after the prototype. When the edge cases show up, the frame budget gets tight, and the build has to work on a machine that isn't mine. That's usually where the interesting engineering is.

Currently

Building Director’s agent surface, and porting models to Candle.

Open to

Engineering roles on inference, tooling or developer products.

01

Systems

Rust, Python, CLI tools, performance profiling, memory and frame budgets.

02

Media

Skia, Vulkan, FFmpeg, layout engines, timeline and animation systems.

03

Product

TypeScript, React, Next.js, Astro, Tauri, local-first architecture.

04

ML

PyTorch to Candle ports, ONNX Runtime, quantization, WASM inference, ComfyUI.

How I work

Build the smallest real version. A prototype that renders one frame is useful. A demo that can't export is not.

Measure before optimizing. Frame times, memory, file size. The number tells you where the work is.

Make failure readable. A good error says what broke, where, and what to try next.

Write it down. If the docs are thin, the project is unfinished.

03 / Contact

Have something that needs to ship?

I’m open to engineering roles and selective contract work.