Neural Style Transfer Playground
Fast style transfer with pre‑baked AdaIN models. Batch stylize in seconds; demo UI supports drag‑drop images, GPU detect, and before/after scrub.
Hey, I'm Milk. Im an AI developer, programmer, and beginner robotics engineer.
14 years old. I love learning new things. I enjoy taking basic ideas and polishing them until they're fun to use.
What I'm into
Model distillation/training tiny LLMs, on‑device inference, creative tools (attempting image rec/vision), and playing video games.
1
years studying AI/ML
0
open repos
2
years of programming
Currently
- Training a keyword spotter that runs fully offline in the browser (WebGPU).
- Prototyping a JS UI for quantized chat models via WebAssembly.
- Porting an MNIST toy to C# + ONNX for Windows desktop.
Featured Work
Some of the builds I'm not embarrassed by. Source where I can share it.
Tiny LLM Chat (Quantized)
Q4_0 weights, streaming tokens, and a clean UI. Runs entirely in the browser. Includes prompt templates, token heatmap, and memory toggle.
Vision: Digit Classifier
Desktop app for drawing digits and seeing logits update live. Quantized ONNX model for instant inference on low‑end hardware.
Audio: Keyword Spotter
Low‑power CNN for on‑device wake word detection. Training notebook + dataset prep + spectrogram augmentations.
Stack & Skills
The tools I actually use.
JavaScript / TypeScript
UI, WASM glue, WebGPU experiments
Python
PyTorch, data pipelines, notebooks
C#
Desktop apps, ONNX interop
ML Ops (lightweight)
Weights & Biases, DVC, Docker
Design / UX
Layout, motion, accessibility
Contact
I'm open to feedback, collabs, and weird ideas. Ping me.
Elsewhere
- GitHub — code and experiments
- X / Twitter — occasional project threads
- Neocities — this site