Aubin Tchoi

Software engineer at Dust, building AI agents and the product around them.

At Dust

Build a Flock, Not a Crowd 

Agents that each keep their own copy of the same know-how drift apart. Skills give that knowledge one home, loaded only when a task needs it, so the whole fleet improves together.

Company Data as a Filesystem 

Our agents kept inventing file paths, so we gave them a filesystem: Notion, Slack, GitHub, and Drive mapped onto one tree they can list, find, read, and search.

Before Dust

3D computer vision at Exwayz, applied mathematics at ENS Paris-Saclay.

An urban point cloud with regularly sampled red keypoints across buildings and the ground.The same scene after NACRE, with ground and flat-wall keypoints removed.

NACRE Detector

Keeps only the keypoints worth matching: edges and corners, scored by the entropy H of nearby normals.

η=H(eig⁡ n n⊤‾)>αln⁡2H(λ)=−∑iλiln⁡λi\begin{aligned} &\eta = H\big(\operatorname{eig}\,\overline{n\,n^{\top}}\big) > \alpha \ln 2 \\ &H(\lambda) = -\textstyle\sum_i \lambda_i \ln \lambda_i \end{aligned}

One LiDAR scan's point-to-plane residuals binned in log space, with the mode and the cutoff marked.

Modal Residual Cutoffs

Sets the ICP outlier cutoff from the mode of point-to-plane log-residuals, which follows the sensor noise and ignores the outliers.

σ^=exp⁡(mode⁡ ln⁡∣r∣)dmax⁡=σ^ 2 erfinv⁡(θ)\begin{aligned} &\hat\sigma = \exp\big(\operatorname{mode}\,\ln\lvert r\rvert\big) \\ &d_{\max} = \hat\sigma\,\sqrt{2}\,\operatorname{erfinv}(\theta) \end{aligned}

3D Semantic Segmentation

Multiscale covariance eigenfeatures, then region growing to smooth labels per object: 75.1% mean IoU.

Ps=voxel⁡(P, 2sr0/5)fs(p)=f(Ps∩B(p, 2sr0))\begin{aligned} &\mathcal{P}_s = \operatorname{voxel}\big(\mathcal{P},\ 2^s r_0 / 5\big) \\ &f_s(p) = f\big(\mathcal{P}_s \cap B(p,\, 2^s r_0)\big) \end{aligned}