a segmentation mask is a pretty fun collision surface.
today’s rainy weather in Cracow felt like a good excuse to try it: on-device person segmentation making rain splash against a silhouette.
React Native ExecuTorch + Vision Camera + Skia.
connecting Vision Camera to React Native ExecuTorch is pretty straightforward: resize a frame, run the model, use the mask 👇
github.com/software-mansion/…
This is what you can build in less than an hour with coding agent + react-native-executorch 🔥
A photo scrubber that finds private text, lets you review the masks, and exports a redacted image. All inference runs locally.
Check out what else is possible in the gallery showcase 👇
github.com/software-mansion-…
I wanted to see how small a browser runtime for ExecuTorch's WebGPU kernels could be, so I built an experimental one in TypeScript.
Unlike the official browser backend, it needs no WASM. It has a minimal core and self-contained kernels, which worked really well for closed-loop agent optimization.
Here's DepthART running in the browser. Still experimental, with lots unsupported.
Trying this makes me appreciate my colleagues' work on ruNNtime at Software Mansion even more. Check out:
docs.swmansion.com/runntime/
Code, if you're curious:
github.com/barhanc/wgpu-exec…
1
94
Try reversing a truck with two trailers. It's a highly unstable system: small steering errors quickly turn into a jackknife.
I trained a transformer to steer it, then built a playable browser simulation using TypeGPU for rendering and ruNNtime for inference.
Playable demo + code in the reply.
Great work on ruNNtime, @_jakubchmura and @_kkaczmar_!
1
2
270
Draw your own route: barhanc.github.io/jackknife/
Code: github.com/barhanc/jackknife
Training uses randomly sampled curves: simulate driving forwards, reverse the trajectories, then add perturbed trailer angles with LQR corrections to teach recovery.
86
