Skip to content

Tech & Tools

What actually makes the walkthrough

Three delivery routes run today. A fourth thing is being built underneath them: an engine that takes a floor plan in and returns a walkable space, without a modeller touching it first.

02

The engine, in development

Status

The engine described below is partially built and is not yet taking outside jobs. It is documented here because it is the direction of the company, not because it is on sale. Everything in section 01 is what you can commission today.

A floor plan goes in. A walkable, furnished, correctly lit space comes out. The hard part is not the rendering, which is well understood. It is reading an architect’s drawing accurately enough that the geometry is true to the millimetre, and knowing which parts the machine got wrong.

So the pipeline is split in two. The first six stages are cheap, fast, and finish with a blueprint a human confirms. Only after that confirmation does anything expensive run. Charging before the human has looked is how you generate something wrong and then refund it.

Pipeline stages
#StageWhat it does
01IntakeClassify the input, grade its quality, reject or warn early
02AnnotationRead dimensions, labels, scale bar and north arrow off the drawing
03GeometryWalls, doors and windows resolved into a vector graph
04CalibrationFuse the read dimensions with the geometry to recover true metric scale
05TopologyRooms, adjacency, watertightness, validation
06ReconciliationBuild the human-facing blueprint, surface what is uncertain, capture the correction
07SemanticsClassify room types, which drives staging and materials
08StagingLay out furniture under collision and reachability constraints
09MaterialsTurn a finishes schedule into PBR assignments and option sets
10LightingSun path from location and orientation, fixtures, bake configuration
11RenderGPU bake and asset generation, cost-capped and resumable
12QualityAutomated checks first, a human queue only for what gets flagged
13PublishCompress, tier the detail, write the manifest, push to CDN, issue the link

Stages 01–06 run before any render cost is committed. Stage 06 is where a person confirms the blueprint, and where their corrections are captured and fed back.

03

Every correction is an asset

When a human fixes a wall the machine misread, that fix is not discarded once the job ships. It is kept, with the drawing that caused it.

That matters commercially rather than just technically. The real cost of running this pipeline is human review minutes, so every point of accuracy gained is review labour removed. The corpus of corrections is the part a competitor cannot copy by buying the same GPUs.

The moat is not the renderer. It is knowing which drawings break it.
04

Choices and why

Orchestration

Durable workflow execution

Thirteen stages that each need to be independently retryable, versioned, and resumable after a crash. A GPU step that dies at minute forty cannot restart the job from zero.

Compute

Rented GPU, spot-first, multi-provider

Generation is bursty. Owning hardware would mean paying for idle silicon between projects, and spot capacity across several providers is both cheaper and harder to be cut off from.

Viewer

glTF with Draco and KTX2, progressively streamed

For authored architectural geometry this is the correct delivery format. Gaussian splatting is a capture technique and does not apply to a building that has not been built.

Delivery

Object storage behind a CDN, immutable cache headers

A walkthrough that takes 147 MB to open on a phone is a walkthrough nobody sees. Compression and cache policy are part of the product, not an optimisation to do later.

If you want the thing that ships today, it ships today.

The engine will be worth talking about when it is finished. Until then the three routes above are real, costed, and have a timeline against them.

Toasty Media Pvt. Ltd. · Bangalore