From algorithm
to physical
object
We build end-to-end systems that bridge artificial intelligence and physical fabrication — from scanning a real-world object to its parametric CAD model, from an evolutionary algorithm to a finished 3D print.
Native Research is where machine intelligence meets the workshop.
We work across the full loop: perception → reasoning → geometry → fabrication. A point cloud becomes a parametric solid; a stress field becomes a lattice; a slicer learns a new kind of printer; an agent runs the experiment overnight.
Everything is built and validated on our own infrastructure and print farm — hypothesis, prototype, real object. We publish what we can as open source.
Reverse engineering
From scans & point clouds to clean, parametric STEP. Section analysis, primitive-free reconstruction, ML-assisted CAD recovery.
Advanced additive mfg
Belt printing, multi-material toolchanging and FEM-driven structural infill — pushing slicers and hardware past their defaults.
Autonomous systems & vision
Evolutionary coding agents, local vision-language perception of design tools, persistent situational-awareness assistants.
AI-piloted CAD
Driving real CAD engines programmatically across a networked farm — sketches, extrudes and assemblies generated and verified by machines.
Robotics & automation
RTK-GPS autonomous machines, home & lab automation, reviving orphaned hardware with open firmware.
Context & tooling
Persistent context graphs, semantic code indexing and the operational glue that lets long-running R&D actually compound.
Triangle meshes (STL / OBJ / 3MF / PLY) into genuine B-Rep solids exported as STEP, on an OpenCASCADE kernel. Topology is shared between faces from the moment they are built, not reconstructed afterwards by a sewing pass — 2.6-3.2x faster than the build-then-sew approach on the same kernel, with identical integrated volume. The default verbatim engine is exact: one planar face per triangle, no fitting, no invented geometry. The trueform engine recovers planes and cylinders analytically, and rebuilds a band of facets as a cylinder only when the measured sagitta shows it came from a tessellator — every doubt resolves to keeping the facets, because a wrong rebuild invents material. Watertightness is reported honestly: an open shell is never dressed up as a solid. Runs in the browser, with select-and-cut editing, or as a CLI.
Belt-printer fork of OrcaSlicer for 45° inclined-belt machines (CR-30, IdeaFormer IR3 V2) — inclined-gantry geometry, gravity-aware supports, and a safety gate every change has to pass before it reaches a real machine. AGPL-3.0, hardware-validated.
Mainline OrcaSlicer with a full parametric, sketch-first CAD environment built into the slicer itself: a real geometric constraint solver on an OpenCASCADE B-rep kernel, the complete feature timeline (extrude, revolve, sweep, loft, pattern, hole, thread, shell, draft, fillet, chamfer, boolean), STEP import as editable B-rep. Design, slice and print in one window — no export, no re-import.
The same in-slicer parametric CAD environment, integrated into Snapmaker Orca and submitted to the Snapmaker U1 Innovation Fund. Open source, AGPL-3.0.
Search 3D models from inside OrcaSlicer and drop one straight onto the plate. Two rules shape it: it never opens a web browser, and every model it offers is downloaded by the plugin itself — so a platform is listed only if its files can be fetched without a login. The licence of every result is shown before anything downloads.
An open catalogue of filament profiles that works for every slicer: pick a filament, generate a tuned profile for the machine and slicer you actually own — OrcaSlicer, Bambu Studio, PrusaSlicer, Cura. Free, community-tested, and the catalogue the OrcaSlicer plugin imports from.
The catalogue, inside the slicer: find a filament profile and import it into OrcaSlicer's filament library as a normal user preset, inheriting from the closest system preset. 514 profiles in the catalogue, no account and no API key.
One click from an Onshape Part Studio to OrcaSlicer on your own machine — a browser extension plus a helper bound to 127.0.0.1. No server, no cloud middleman, no API keys: it uses the Onshape session you are already logged into.
Autonomous evolutionary coding system for PLY to STEP reconstruction via section analysis.
Partitioning large models across a heterogeneous print farm to minimise makespan.
Stress-aware lattice infill — FEM fields drive strut radius for strength-to-weight.
Local VLM perception layer feeding machine-parsable descriptions of GUI design apps.
3D scans (STL/OBJ/PLY) to parametric STEP using ML and CadQuery.
Native multi-filament changer driver for Klipper via the Happy Hare framework.
Real-time 3D scan to NURBS/STEP via projector-camera structured-light feedback.
mesh2step ships, and the OrcaSlicer family grows
mesh2step is public and live in the browser: triangle meshes to genuine B-Rep STEP solids on an OpenCASCADE kernel, with topology shared at construction time instead of reconstructed by a sewing pass — 2.6-3.2x faster than the build-then-sew approach, at identical integrated volume. Around it, the slicer work has grown into a family: ShidaoSlicer for 45° belt machines, Orca-Cad and SnapOrca Design bringing parametric sketch-first CAD inside the slicer window, plus FilamentProfiles.Org and three plugins (model search, filament profiles, Onshape one-click). Every published project now links straight to its source.
Native Research website goes live
First public one-pager deployed to Netsons hosting and served over HTTP. Dark maker/hacker theme on the brand palette, with research tracks and a live project index. DNS A record for the apex added; AutoSSL pending certificate issuance.
Autonomous content pipeline online
The site now rebuilds and republishes itself. A daily job collects project manifests and session log entries, regenerates the page, and deploys it to the hosting docroot over FTPS — no manual steps. HTTPS is live with an issued certificate.
Open by default
Research and tooling published as open source wherever licensing allows, so others can build on it.
Local-first & Linux
Self-hosted compute, on-prem models and a fully Linux stack — data and experiments stay in-house.
Algorithm to atom
Nothing counts until it exists as a real object on the bench. The print is the proof.