Consumer scan apps snap everything to a template and hide the real deviation. Enterprise survey tools are accurate but still draft by hand on the desktop. Fazzad streams the point cloud, extracts the geometry from measured evidence, and asks a person only about the few spots it can't prove — automating most of a survey-grade DWG without giving up defensible dimensions.
We don't hallucinate a plan — we transform evidence. Multi-band Z-slices isolate wall returns from floor clutter, normal consensus filters co-planar surfaces, and the wallnet graph solver outputs survey-grade DWG geometry with a per-wall QA audit log.
The same pipeline shown above, in plain steps — what happens at each stage, where the AI helps, and what you get. The geometry is measured from the scan, not imagined by a language model; the AI is a layer of judgment on top, and it never signs off alone.
Upload the point cloud (LAS / LAZ / E57). It streams as COPC, and a slice at wall height is cut out of it.
WHYA full scan is billions of points of everything — furniture, ceilings, people. A thin band at wall height, with the clutter gone, leaves only what defines the plan, so every later step reasons about walls, not noise.
AIStrips furniture, fixtures and clutter, so only the building structure moves forward.
The slice is projected straight down into an ortho image, then counted into a density grid — walls show up as bright bands, loose clutter stays faint.
WHYRaw 3D points are hard to draw from. Flattening to a scaled top-down picture turns “where are the points” into “where are the walls” — a clean 2D image the next step can trace, at true scale.
no AIPure measured geometry — nothing is inferred at this stage.
The mask is traced into wall lines, paired up by thickness, and solved into a network — corners typed L / T / X, openings located.
WHYA CAD drawing isn’t pixels — it’s connected lines with real topology. Solving the network is what makes the output editable, buildable CAD, not just a picture of a plan.
AI · ANALYSTReviews the geometry against the room layout, repairs near-miss junctions, and proposes corrections a raster can’t see.
Every wall gets a status. Walls the scan strongly supports verify on their own; the rest stop and ask — each with its scan evidence attached.
WHYA scan is never perfect — occlusions, furniture against walls, ambiguous gaps. Rather than guess and hide the error the way template apps do, we surface exactly what we can’t prove, so the drawing stays honest and measured.
AI · CRITIC + VISIONA second model tries to refute every call — only survivors verify. Vision reads the scan photo at uncertain gaps: doorway or occlusion? window or reflection? Anything left unproven becomes a human flag.
Approved geometry is snapped to measured thickness, dimensioned, put on AIA layers, and written to a signed DWG plus PDF sheets and a per-wall QA record.
WHYThe customer needs a survey-grade file they can build and permit from — not a sketch. True thickness, dimensions and layers make it that standard DWG, and the QA record keeps the accuracy defensible.
AI · JUDGERuns on every job — names rooms from tagged site photos, reconciles each room’s area across the wall-cycle and its measured dimensions, and logs a signal that makes the next plan better.
AI never draws the walls — measured evidence does, which is why the dimensions hold up in due diligence.