Shltr on Vultr OS

After a flood, families have to prove every loss to get aid, and in the same week scammers target them with fake relief websites. Professionals document damage with 3D digital twins (Matterport reports claim estimates 400% faster)

Sheltr gives the survivor that evidence from one phone photo or video. An agent on Vultr (glm-5.3 on Serverless Inference) plans the case and dispatches every risky step to its own Microsandbox microVM on a Vultr VX1:

it rebuilds the room in 3D (Depth Anything V2, no network), finds and prices the damage, writes Python to measure it (shown live with its retry loop), builds a 4-page PDF with SHA-256 fingerprints, opens a suspicious "aid" link in a sealed browser, where the forced malware download is captured and the data theft is blocked, fills the aid application live on screen.

The survivor can edit every field the agent typed. Nothing is submitted until they approve, and the submit runs in a fresh microVM. Keys never enter a sandbox, the sandbox host has no public ports, and every microVM is destroyed after its job. The aid portal and scam site are our own mocks; the agent never touches a real government site.

Note: the demo URL is substituted with the GitHub repository link.