No. 12

From model and site data to a place you can walk

Context
Personal R&D
Year
2026
Type
Research and development
Tools
Revit, Houdini, Blender, USD, Unreal Engine, Python, Layer

Turning site data, GIS and building models into a model a client can walk through usually means a great deal of modeling by hand. This pipeline computes it instead, so a change is a rule and a re-run.

I designed the pipeline, decided what it should do, set the checks and judged the results. AI coding agents wrote most of the code and ran the builds under my direction. I say that plainly because directing that kind of work is the skill I want to bring to a firm.

The same pipeline produced a design study for a playground in this park: Rainbow Park playground.

I set out the idea behind this work in One database across the building life cycle.

01

What comes out

02

This is what the pipeline is for: a place you can look at and walk through.

From above the clouds to the shore of Zorinsky Lake, in the Unreal viewer. A scripted tour, captured as it played.
ClipFrom above the clouds to the shore of Zorinsky Lake, in the Unreal viewer. A scripted tour, captured as it played.
Zorinsky Lake from high above in the viewer: the long, branching lake with wooded shores, a large lawn below the dam at the bottom, a boat ramp and parking at the lower right, and suburban streets and houses all around.
Zorinsky Lake from above in the viewer, built from public data.

02

What runs with nobody at the controls

01

The headless steps Two lanes, read left to right. Building lane: a Revit, IFC, Rhino or SketchUp model goes to Houdini, run headless (hython) as versioned tools, which reads it in, works out the room of each element, adds analysis (sun hours, view out, distance to the nearest exit) and details (wall trim, stair finishes, wall layers), and writes USD with each element's id as an attribute (bim:unique_id). Site lane: public data (elevation, lidar, aerial imagery, OpenStreetMap, parcels, lake depths) and a rules file, site.json, go to the hub, headless Python with no 3D program open, which builds terrain, ground classes, paths, buildings with door records, drives and walks, furniture, tree placement and design options, and writes USD with ids. A side input to the hub is the asset library: Houdini measures each library tree and renders its distance cards; Blender is the library and look host. Both lanes end in the Unreal level, where every element is its own actor carrying its id as a tag and a row in elements.json. What this unlocks: change a rule and re-run. The garage rule changed the garage wall of 709 houses; yard plans read the hub's door records, with 0 disagreements over 4,768 houses; 1,700 drives and 856 walks were drawn from the hub's door records. INPUT HEADLESS STEP FORMAT PLAY BUILDING MODELS Revit, IFC, Rhino or SketchUp Houdini run headless (hython) as versioned tools read in · rooms analysis: sun, view, exit distance details: trim, stairs, wall layers USD ids as attributes bim:unique_id SITE Public data elevation, lidar, aerial imagery, OpenStreetMap, parcels, lake depths Rules file site.json: classes, overrides, options, tree palette SIDE INPUT Asset library Houdini measures each tree and renders its distance cards; Blender: library and look host The hub headless Python, no 3D program open terrain · ground classes · paths buildings with door records drives and walks · furniture tree placement · design options USD ids as attributes bim:unique_id Unreal level every element its own actor, carrying its id as a tag and a row in elements.json WHAT THIS UNLOCKS: CHANGE A RULE, RE-RUN 709 houses changed garage wall when the garage rule changed 0 disagreements between door records and yard plans, over 4,768 houses 1,700 drives, 856 walks drawn from the hub's door records
The headless steps Stacked version. Building models: a Revit, IFC, Rhino or SketchUp model goes to Houdini, run headless (hython) as versioned tools: read in, rooms, analysis (sun, view, exit distance) and details (trim, stairs, wall layers), written to USD with each element's id as an attribute (bim:unique_id). Site: public data (elevation, lidar, aerial imagery, OpenStreetMap, parcels, lake depths), a rules file (site.json) and, as a side input, the asset library (Houdini measures each tree and renders its distance cards; Blender is the library and look host) go to the hub, headless Python with no 3D program open: terrain, ground classes, paths, buildings with door records, drives and walks, furniture, tree placement and design options, written to USD with ids. Both end in the Unreal level, every element its own actor carrying its id as a tag and a row in elements.json. What this unlocks: change a rule, re-run. 709 houses changed garage wall when the garage rule changed; 0 disagreements between door records and yard plans over 4,768 houses; 1,700 drives and 856 walks drawn from the hub's door records. BUILDING MODELS Revit, IFC, Rhino or SketchUp Houdini run headless (hython) as versioned tools read in · rooms analysis: sun, view, exit distance details: trim, stairs, wall layers USD ids as attributes, bim:unique_id SITE Public data elevation, lidar, aerial imagery, OpenStreetMap, parcels, lake depths Rules file site.json: classes, overrides, options, tree palette SIDE INPUT Asset library Houdini measures each tree and renders its distance cards; Blender is the library and look host The hub headless Python, no 3D program open terrain · ground classes · paths buildings with door records drives and walks · furniture tree placement · design options USD ids as attributes, bim:unique_id PLAY Unreal level every element its own actor, carrying its id as a tag and a row in elements.json WHAT THIS UNLOCKS: CHANGE A RULE, RE-RUN 709 houses changed garage wall (garage rule) 0 disagreements doors against yards over 4,768 houses 1,700 drives, 856 walks from the doors
The hub is Python with no 3D program open. It builds the terrain, eight ground classes, paths from a distance field, buildings massed from lidar with a record of their doors, drives, walks, furniture, tree placement and design options. Houdini runs headless on building models: it reads them in, works out which room each element stands in, computes sun hours, view out and the walk to the nearest exit, and adds trim, stair finishes and wall layers. For the park it only measures the library trees, renders their distance cards and converts library assets to USD.

03

A change is a re-run

02

A garage rule put each garage door on the wall facing the street. On the re-run, 709 houses changed garage wall and 366 gained doors. The hub writes each building's doors into its record, and the yard plan reads those records instead of running the rule again. Over 4,768 houses there are 0 disagreements. 1,700 concrete drives and 856 walks were drawn from the doors the hub planned. A meadow outline and species weights added to the rules file took the site from 36,797 to 48,470 grass clump points in one rebuild.

Two views of the marina parking lot in the park model from the same camera and light. Left, existing: an open asphalt lot. Right, proposed: the same lot with a planting island of trees and flowers.
Existing against proposed at the marina lot, from the same camera and light. The planting is a design option: data in the rules file, built as its own model and switched in the viewer.
Switching the marina lot between the existing site and the proposed planting, in the viewer's Controls panel.
ClipSwitching the marina lot between the existing site and the proposed planting, in the viewer's Controls panel.

04

One id, end to end

04

Every element keeps one id from the source file to the viewer. The id lives in an attribute, never in a name or a point number, which Houdini renumbers freely. Reading the Revit file without Revit matched all 11,959 unique ids, with geometry within 5 mm at 99.4 percent on one model and 94.7 percent on the other.

One id, end to end One element followed through the chain. A building element, live in the school model: its Revit UniqueId, read without Revit (11,959 of 11,959 exact), becomes the Houdini attribute bim:unique_id (plus sem:room and the ana:* analysis channels), stays bim:unique_id on its USD prim (the prim's path is derived from the id), becomes an Unreal actor tag and a row in elements.json; clicking it shows the inspector, filters and a note keyed to the id, and opens the element's record in Layer. A park bench at Lake Zorinsky, id feature:bench/furn-f096/bench: item F096, the hub's feature id, USD bim:unique_id with sem:class furniture and ifc:class IfcFurniture, an Unreal actor tag and data row, the inspector (Unique ID, category, class; Isolate, Same type). The last two links, opening its Site Furniture row in Layer and the notes push to Layer, are not yet live. What the id drives: select and inspect; filter (isolate, ghost others, same type, family or category); color by data in the building model; pin a note to an element; existing against proposed; schedules in Layer keyed by the same ids. SOURCE HEADLESS STEP USD UNREAL IN THE VIEWER RECORD A BUILDING ELEMENT · LIVE IN THE SCHOOL MODEL opens its record Revit UniqueId read without Revit: 11,959 of 11,959 exact Houdini bim:unique_id plus sem:room and the ana:* channels USD bim:unique_id on the element's prim; its path is derived from the id Unreal bim:unique_id=… an actor tag, and a row in elements.json Click inspector, filters, a note keyed to the id Layer the element's record A PARK BENCH · LAKE ZORINSKY id feature:bench/furn-f096/bench opens its record notes push Item F096 the number the feature id carries The hub makes the feature id and writes it to USD USD bim:unique_id sem:class furniture ifc:class IfcFurniture Unreal an actor tag, and a row in elements.json Inspector Unique ID, category, class; Isolate, Same type; a Ctrl N note keyed to the id Layer its Site Furniture row NOT YET LIVE WHAT THE ID DRIVES Select and inspect click an element, read its data Filter isolate, ghost others, same type, family or category Color by data in the building model Pin a note Ctrl N keys the note to the element's id Existing against proposed options carried by saved views, notes and meeting stops Schedules rows in Layer keyed by the same ids
One id, end to end Stacked version. A building element, live in the school model: Revit UniqueId, read without Revit (11,959 of 11,959 exact); Houdini attribute bim:unique_id, plus sem:room and the ana:* channels; USD bim:unique_id on the element's prim; Unreal actor tag and a row in elements.json; click: inspector, filters, a note keyed to the id; its record in Layer. A park bench at Lake Zorinsky, id feature:bench/furn-f096/bench: item F096; the hub's feature id; USD bim:unique_id with sem:class furniture and ifc:class IfcFurniture; Unreal actor tag and data row; the inspector (Unique ID, category, class; Isolate, Same type; a Ctrl N note keyed to the id). Opening its Site Furniture row in Layer and the notes push are not yet live. What the id drives: select and inspect; filter; color by data in the building model; pin a note; existing against proposed; schedules in Layer keyed by the same ids. A BUILDING ELEMENT · LIVE in the school model Revit UniqueId read without Revit: 11,959 of 11,959 exact Houdini bim:unique_id plus sem:room and the ana:* channels USD bim:unique_id on the prim; its path is derived from the id Unreal bim:unique_id=… an actor tag, and a row in elements.json Click inspector, filters, a note keyed to the id opens its record Layer the element's record A PARK BENCH · LAKE ZORINSKY id feature:bench/furn-f096/bench Item F096 the number the feature id carries The hub makes the feature id and writes it to USD USD bim:unique_id sem:class furniture, ifc:class IfcFurniture Unreal an actor tag, and a row in elements.json Inspector Unique ID, category, class; Isolate, Same type; a Ctrl N note keyed to the id opens its record · notes push NOT YET LIVE Layer its Site Furniture row WHAT THE ID DRIVES Select and inspect Filter isolate, ghost, same type or category Color by data building model Pin a note Ctrl N, keyed to the id Existing against proposed Schedules in Layer, same ids
In the viewer the id selects an element, filters by type, family or category, and keys a note pinned with Ctrl N. In the school model, clicking an element opens its record in Layer. On the park bench, that last link and the notes push are not live yet.
The park model in the viewer with a bench selected beside a path. The inspector panel on the right shows Bench (F096) bench, Unique ID feature:bench/furn-f096/bench, Category Furniture, Class furniture, IFC class IfcFurniture, and the buttons Isolate, Ghost others, Same type, Same family and Same category.
A bench selected in the park model. The inspector shows its Unique ID, feature:bench/furn-f096/bench, with its category, class and IFC class, and the filters.
Clicking a bench opens the Inspector with its Unique ID.
ClipClicking a bench opens the Inspector with its Unique ID.
In the school model, clicking an element opens its record in Layer. From a screen recording of the school model, October 2026.
ClipIn the school model, clicking an element opens its record in Layer. From a screen recording of the school model, October 2026.

05

A place you can use

06

The model is not a film. You fly, walk, paddle, change the hour, the month and the weather, select things and leave notes. The interface is on screen in these clips on purpose.

The time of day stepped from morning to the low sun, from the dam crest. The hour shows in the status bar at the top.
ClipThe time of day stepped from morning to the low sun, from the dam crest. The hour shows in the status bar at the top.
The month stepped from August toward winter; the trees turn.
ClipThe month stepped from August toward winter; the trees turn.
Sunset over the lake with the Controls panel open on Environment, then the weather set to snow. My own screen recording.
ClipSunset over the lake with the Controls panel open on Environment, then the weather set to snow. My own screen recording.
Paddling a kayak out from the boat ramp, in third person.
ClipPaddling a kayak out from the boat ramp, in third person.
Walking the dam path, then switching to the bike. From the same recording.
ClipWalking the dam path, then switching to the bike. From the same recording.
The pier and its shelter at golden hour; the Controls panel opens on the key map. From the same recording.
ClipThe pier and its shelter at golden hour; the Controls panel opens on the key map. From the same recording.

06

The school model in the viewer

06

Walking a floor in third person. From a screen recording of the school model, October 2026.
ClipWalking a floor in third person. From a screen recording of the school model, October 2026.
The same floor at eye level, in first person.
ClipThe same floor at eye level, in first person.
The building pulled apart by level, each level labelled.
ClipThe building pulled apart by level, each level labelled.
One level at a time, as a plan cut.
ClipOne level at a time, as a plan cut.
Searching rooms and levels; choosing a room flies there.
ClipSearching rooms and levels; choosing a room flies there.
Stepping the time of day on the facade.
ClipStepping the time of day on the facade.

07

Checked against the place

02

I photographed the park on the morning of 31 October 2026. 258 of the 328 photographs got a matched camera in the model; 6 of them are refined so far, against a target of 40. The viewer can wipe between a photograph and the model from the same camera.

Side by side: a photograph of the park's marina parking lot under an overcast sky, with parked cars, a flagpole, a shelter and a tree line, and the model of the same lot from the same camera, with its shelter, tree line and lawn.
My photograph of the marina lot, left, and the model from the same camera, right.
The viewer's Compare wipe at the marina lot: the photograph on the left of a vertical line, the model on the right of it.
The viewer's Compare wipe: the photograph left of the line, the model right of it, from the same camera.

08

How the agents are run

01

The tooling lives in four repositories on GitHub. The engine alone has 504 commits in five weeks, from 11 October to 15 November 2026, and 419 of them carry an AI co-author line. The sessions were Claude models, with ChatGPT/Codex as the builder for four blocks before 7 November. Before any work, an agent reads a fixed set of start files: the engine's rules, the map of the part it is working in, which links that part's contracts, and my general working agreements. There are 19 contracts, 14 of them for the spatial pipeline. Three of the rules: "Proof before claims. A still you looked at, a number from a log, a hash. If a check could not run, say so." "Push or publish only when Josh authorizes it." "Decide and go. When a choice is Josh's but he is away, take the recommendation written in the brief … record the decision in your log under 'Decisions for Josh', and continue." Work is cut into blocks. One session orchestrates: it writes a brief, hands it to a single worker session, checks what comes back and keeps the log. In the park's log, 29 blocks ran this way, between 7 and 15 November 2026, with a pause from the evening of 7 November to 13 November. 60 briefs are on file, some from earlier blocks. Nothing counts until it is checked. The engine's automated check has 70 numbered sections, covering things like hand work surviving every scene sync, parts within ±2 mm of their declared bounds, and a license gate that blocks any asset that is not CC0, public domain or owned until its license is cleared. Then the orchestrator opens the final stills at full size and looks. Its notes include "Not as reported: stills/after/ac-gulls.png shows no gull." One check exists because of a failure. On 23 October 2026 an index run on one folder shrank the library's id table from 4,662 assets to 55. It was recovered from sidecars, and guards were added so that "a table that falls below half its last size is refused." That guard is check section 57. I keep the decisions: the order of work, licenses, deletions and every push to GitHub. Since 13 November 2026, on my word, a block may delete its own superseded attempts. It is not tidy. During the 29 park blocks the engine's full check was not run per block, and its latest recorded full quick run, on 22 and 23 October, had 352 passed and 3 failed. Workers' reports were wrong more than once; those errors are caught only because the orchestrator re-checks the reports. 15 of the 29 block headings say "LANDED" rather than "COMPLETE", and one more says "COMPLETE with open items". And it all runs on one machine, one heavy job at a time.

One block, start to finish A loop, read clockwise. 01, Josh sets the order: in chat, recorded in BUILDER-LOG.md ("Josh changed the order", "Yes, quality next"); later his call: pushes, deletions, licenses. 02, the brief: BRIEF-tag.md, written by the orchestrator with Josh's words, the verified state with paths, numbered tasks and the proof list, bound to the 12 standing rules (HANDOFF-BUILDER.md section 2), AGENTS.md and the contracts, spatial 01 to 14. 03, one worker builds: one Opus worker per block; rules in site.json; the hub, site-hub.mjs build, release, verify-export; Houdini or Unreal jobs through unreal.mjs and run.mjs; then a cook; every heavy launch through heavy.mjs, a lock file, logged in launches.jsonl. 04, automated checks: test_site_hub.py 100/100, rule unit tests, the brand scan, the launch proof (0 Error lines, LoadMap time), the Model SHA-256 in three control files and a census of actors and instances; engine-wide, spatial/check.mjs has 70 sections, its last full quick run 352 passed, 3 failed. 05, the orchestrator looks: it opens the final stills and the photo-against-model sheets at full size, measures color boxes, and writes "Seen by the orchestrator" and "Not right, open", for example "Not as reported: ... shows no gull". 06, the record: README-tag.md, a CHANGELOG line, an engine commit by file name on spatial, the BUILDER-LOG.md entry, the HANDOFF-BUILDER.md section 0 update and the Layer Builds row. Open items go into the next brief, or are carried as "Not right, open", back to 02. To the side, a gate: GitHub, only on Josh's word: a git push of spatial, code and fixtures only, never automatic. ONE BLOCK · READ CLOCKWISE 01 · JOSH 02 · BRIEF 03 · BUILD 04 · CHECK 05 · LOOK 06 · RECORD NEXT BLOCK Josh sets the order in chat, recorded in BUILDER-LOG.md “Josh changed the order” “Yes, quality next” later his call: pushes, deletions, licenses The brief BRIEF-<tag>.md by the orchestrator: Josh's words, verified state with paths, numbered tasks, the proof list bound to the 12 standing rules (HANDOFF-BUILDER.md §2), AGENTS.md and the contracts, spatial 01 to 14 One worker builds one Opus worker per block rules in site.json the hub: site-hub.mjs build, release, verify-export Houdini or Unreal jobs: unreal.mjs, run.mjs; then a cook every heavy launch through heavy.mjs: a lock file, logged in launches.jsonl Automated checks test_site_hub.py 100/100 rule unit tests · brand scan launch proof: 0 Error lines, LoadMap time Model SHA-256 in three control files a census of actors and instances engine-wide: spatial/check.mjs, 70 sections; last full quick run 352 passed, 3 failed The orchestrator looks opens the final stills and the photo- against-model sheets at full size, measures color boxes writes “Seen by the orchestrator” and “Not right, open” “Not as reported: … shows no gull” The record README-<tag>.md, a CHANGELOG line an engine commit by file name on spatial the BUILDER-LOG.md entry the HANDOFF-BUILDER.md §0 update the Layer “Builds” row Open items go into the next brief, or are carried as “Not right, open” GitHub: only on Josh's word a git push of spatial, code and fixtures only; never automatic
One block, start to finish Stacked version, read down, then back up. 01, Josh sets the order: in chat, recorded in BUILDER-LOG.md ("Josh changed the order", "Yes, quality next"); later his call: pushes, deletions, licenses. 02, the brief: BRIEF-tag.md, by the orchestrator: Josh's words, the verified state with paths, numbered tasks, the proof list; bound to the 12 standing rules (HANDOFF-BUILDER.md section 2), AGENTS.md and the contracts, spatial 01 to 14. 03, one worker builds: one Opus worker per block; rules in site.json; the hub, site-hub.mjs build, release, verify-export; Houdini or Unreal jobs through unreal.mjs and run.mjs; then a cook; every heavy launch through heavy.mjs, a lock file, logged in launches.jsonl. 04, automated checks: test_site_hub.py 100/100, rule unit tests, brand scan, launch proof (0 Error lines, LoadMap time), the Model SHA-256 in three control files, a census of actors and instances; engine-wide, spatial/check.mjs, 70 sections, last full quick run 352 passed, 3 failed. 05, the orchestrator looks: the final stills and the photo-against-model sheets at full size, color boxes measured; it writes "Seen by the orchestrator" and "Not right, open", for example "Not as reported: ... shows no gull". 06, the record: README-tag.md, a CHANGELOG line, an engine commit by file name on spatial, the BUILDER-LOG.md entry, the HANDOFF-BUILDER.md section 0 update, the Layer Builds row. To the side, a gate: GitHub, only on Josh's word: a git push of spatial, code and fixtures only, never automatic. Open items go into the next brief, or are carried as "Not right, open", back up to 02. ONE BLOCK · DOWN, THEN BACK UP 01 · JOSH Josh sets the order in chat, recorded in BUILDER-LOG.md “Josh changed the order”, “Yes, quality next” later his call: pushes, deletions, licenses 02 · BRIEF The brief BRIEF-<tag>.md by the orchestrator: Josh's words, verified state with paths, numbered tasks, the proof list bound to the 12 standing rules (HANDOFF-BUILDER.md §2), AGENTS.md and the contracts, spatial 01 to 14 03 · BUILD One worker builds one Opus worker per block rules in site.json the hub: site-hub.mjs build, release, verify-export; Houdini or Unreal jobs: unreal.mjs, run.mjs; then a cook every heavy launch through heavy.mjs: a lock file, logged in launches.jsonl 04 · CHECK Automated checks test_site_hub.py 100/100 rule unit tests · brand scan launch proof: 0 Error lines, LoadMap time Model SHA-256 in three control files a census of actors and instances engine-wide: spatial/check.mjs, 70 sections; last full quick run 352 passed, 3 failed 05 · LOOK The orchestrator looks opens the final stills and the photo-against- model sheets at full size, measures color boxes writes “Seen by the orchestrator” and “Not right, open” “Not as reported: … shows no gull” 06 · RECORD The record README-<tag>.md, a CHANGELOG line an engine commit by file name on spatial the BUILDER-LOG.md entry the HANDOFF-BUILDER.md §0 update the Layer “Builds” row GitHub: only on Josh's word a git push of spatial, code and fixtures only; never automatic NEXT BLOCK Open items go into the next brief, or are carried as “Not right, open”: back up to the brief
One block, from my order to the record. Open items feed the next brief; a push to GitHub happens only on my word.

09

What is not done

Clicking through to Layer is not set up on the park model, and sending notes to Layer has only run as a dry run. Only 6 of the 40 target cameras are refined. The camera moves ease to a stop between tour legs, and the time of day changes in half-hour steps; smooth versions need engine work. The houses are massed by rule from lidar over OpenStreetMap footprints, not surveyed. The tree species are illustrative. The data is visualization grade, from public sources. It is not a survey. What I want to find out next is how this holds up on a live project, with a real team's models and deadlines.