Research note. By Sista Seetaram
Notes from building BuildBridge Studio, Setu's first product, May to October 2026: a 2D floor plan becomes a checked 3D model and then a render, and a person approves every stage. What worked, what failed, and the numbers.
Built. May to September 2026.
Every early render looked right and was wrong, so I stopped chasing better pictures and built the geometry first, where it can be checked.
I began with an AI agent meant to carry an Indian interior project from the first brief to handover. By 26 May it had a database, a mood board and palette stage, Vaastu (traditional Indian layout rules) and climate notes, and material and furniture sheets for a test flat, backed by 32 automated tests.
The first renders looked finished and broke the plan. Sofas flipped, doors moved, an open dining area was sealed shut. I tried seven ways to make an image model respect the drawing, from coordinate-heavy prompts to grids and plain grey 3D block guides. All seven failed, and on 27 May I paused rendering.
The plan is a contract: style may change, layout may not, unless the designer approves.
A vision model reads the plan into a coordinate file in feet. The Three.js library builds a whitebox from it, a plain grey 3D model. Checks run before every paid step and stop on any error. On 12 June I moved the reader from Claude Sonnet 4.6 to Gemini 3.1 Pro. A stronger model still cannot read detail the image does not hold.
I hand-labelled 99 boxes on the test plan so that scores meant something. Giving models the plan's symbol key made Gemini's detection worse and did not help the others. A scorer bug that swapped one model's x and y coordinates had skewed the leaderboard. On 21 June plain OpenCV, with no model, located 22 of 38 walls, but its settings were tuned to this one plan's drawing style, so I did not build on it.
On 16 June one recipe held the layout: a line drawing projected from the whitebox, plus a detailed prompt, into an image model. Looking east, you cannot see the prayer room, so prompts now name only what the camera sees. Picking the best of three drawings raised edge coverage, the share of the whitebox's lines that survive, from 0.62 to 0.87 (one camera, 23 June). A MistoLine model I built and deployed rendered furniture as blocks, so I dropped it.
Intake
Concept
Extract
Verify
Fixtures fix
Whitebox
Line-art
Render
Refine
PackageFive separate lines of experiments were no use to a designer. A status check found four stages showed screens but did no work. I wired them up and joined everything into one rail. On 3 July four AI agents with set roles, including a metric sceptic, argued the architecture; I took their verdict: repair, do not rebuild.
Walls the model invented survived three correction rounds, so I rebuilt walls as 31 explicit line segments from the corrected plan; none appear in the open living area. A test with rules written before any spend put four image models on three cameras. I waived its blind scoring; on my judgement nano-banana-pro won all three, and rendering straight from the shaded whitebox became the default.
On 9 July the pipeline ran end to end on a real office plan, about $0.70 in model calls (estimate, list prices). Then I turned it into a tool a designer can drive: door editing, site measurements, plain-language geometry edits, and refining a render with reference images. Independent AI reviewer agents checked each test-first piece; one caught a script-injection bypass the building agent missed.
A real office room exposed what the test flat had hidden: furniture drawn 18.6% too wide on one axis, a camera placed inside a wall, and prompt bugs I had blamed on the models. The architect and their client accepted one room's render after seven refine passes; an eighth paid call failed. On 12 September all the work merged into one codebase, with 1,281 tests passing.
Renders now go to nano-banana-pro first, with GPT Image 2 as the alternate, and there is no third model. I tested two models that trace the control image's edges: four calls, about $0.11. flux-canny-pro kept the lines and failed as interior design; Z-Image Turbo did not even hold the camera. Holding geometry is necessary and nowhere near sufficient.
Plans as of 11 October 2026.
So far the pipeline has run end to end on one test flat and one real office room, with a person correcting it at each stage. Next, its accuracy has to be measured on plans it was not tuned on.
Almost every number on this page comes from one plan, so plans the pipeline has never seen are the test, counting the correction rounds each one needs. Cropping to a single bedroom found more of its 8 fixtures by name, 3 against 2, while precision fell from 0.67 to 0.43. Cropping is not yet in the live pipeline.
Three attempts to move one piece of furniture, by prompt and by edit, all failed, so the next step puts real furniture shapes, not plain boxes, into the whitebox before anything is rendered.
The first, the general arrangement plan, is the first to build.
A contractor cannot build from a render. I specified a ten-sheet construction drawing set: plans, ceilings, electrical, plumbing, elevations and schedules, each to be drawn from the same verified geometry. The research behind it checked 25 of 120 claims from 26 sources against counter-evidence and rejected 7; three independent designs were critiqued and merged into one specification. No sheet is built.
Today one vision model reads the whole plan. The proposed second version pairs plain image processing for walls with a small trained segmentation network, a U-Net, and stores walls as a connected graph. A licence survey of five candidates on 19 June 2026 found none with downloadable weights for raster floor plans under a permissive licence; CubiCasa5k, for example, is non-commercial only. So training my own model on labelled Indian plans is a gated option, for when volume or error rate justifies it.
I want an editor where a designer changes walls and rooms directly, as geometry, not as pixels in a render. The catch: my coordinate file stores rooms as plain rectangles, with no record of how walls join, so it needs a new file format first. No code exists yet.
The studio runs ten stages, and nothing advances until a person says so.
| # | Stage | What it does | Model or library | Where the person checks |
|---|---|---|---|---|
| 01 | Intake | Captures the brief in chat; turns a PDF plan into an image. | Claude Sonnet 4.6 via OpenRouter; pdftoppm | Approves the brief. |
| 02 | Concept | Mood board, palette, Vaastu and climate notes, carried into every render prompt. | Claude Sonnet 4.6; plain Python | Approves palette and mood. |
| 03 | Extract | Reads rooms, walls, doors, windows and fixtures into one coordinate file in feet. | Gemini 3.1 Pro via OpenRouter; one call, no retry | Approves the paid call. |
| 04 | Verify | Blocking spatial checks: fixtures inside rooms, doors reachable and clear, outline sealed, no voids, walls thin and inside the outline. | Plain Python, no model, no cost | A failed check stops the rail and says how to fix it. |
| 05 | Fixtures fix | Move, resize and rotate furniture, edit doors and windows, enter site measurements. | Konva canvas, scaled from the detected building outline | Saves the corrected plan as one verified record. |
| 06 | Whitebox | A plain 3D model built from the file. Plain-language edits change the coordinate file, never the 3D model directly. | Three.js; Gemini 2.5 Flash for edits | Approves the geometry. |
| 07 | Line-art | A two-point perspective line drawing as a control image. One of two render paths; now optional. | Three.js line projection, then an image model | Picks the cameras. |
| 08 | Render | Shaded whitebox to photoreal in one step. Code builds the prompt in labelled sections and gives directions as left, right and ahead, because image models have no sense of compass directions. | nano-banana-pro (default) or GPT Image 2 via Replicate. A local Mask2Former reject gate runs only when several candidates are requested; off by default. | Approves spending; accepts or rejects each render. |
| 09 | Refine | Edits the accepted render from notes and up to five reference images, keeping its proportions. | nano-banana-pro edit; optional Gemini 3.1 Pro prompt composer | Reviews each revision before it is paid for. |
| 10 | Package | Sends schedules and renders to the shared drive with a single-sheet DXF of the verified plan. | Google Drive and Sheets APIs; ezdxf | Confirms the publish. |
Nearly all of these come from one plan or one room. Read them as early signals, not benchmarks.
0.628
0.302
These two come from different runs with different inputs and scoring, so they are not directly comparable. I show both because the gap is the finding: numbers from one plan move a lot.
0.884, then 0.318
12 of 12, then about 4 of 6
16 of 16 rooms, 36 of 36 items
36 and 37 of 41
274 commits over 124 days
One person directing AI coding agents: 239 of the 274 commits carry an AI co-author line. 47 days with commits, 21 May to 22 September 2026. Counted with git on 12 October 2026.
20 to 1,281 automated tests
25 May to 12 September 2026, from session notes; 1,281 passing at the 12 September merge.
99 hand-labelled boxes
Pixel ground truth on the test plan, 19 June 2026.
20 experiment reports
15 to 23 June 2026. 8 vision models and 33 placement configurations ranked on one plan, 18 and 19 June 2026.
At least 9 multi-agent sessions
Research or review sessions, 18 June to 16 September 2026, from wall research by 12 parallel agents to a four-agent backlog audit.
9 decision records, 13 handoff notes
Files in the project's planning folder, counted 12 October 2026.
88 research notes
In a wiki started 23 June 2026 (97 files with templates, index and log), counted 12 October 2026. Among them, 15 written bets, each with a test that could prove it wrong; one retired, one downgraded, as of 20 September 2026.
26 paid calls, about $2.55
Estimate. One real project, intake to one accepted room render, early September to 12 September 2026, including first renders of three rooms. Estimated from list prices; chat models on subscription excluded.
Commits by month, 2026.
The 26 paid calls behind one accepted room render. The outlined square is the failed call.
An audit on 13 September listed 46 open items and found about 25 that my own notes called pending but were already built.
Direction, an aim not yet reached: from one floor plan to a render you can trust and drawings a contractor can build from, with a person approving every step.
Setu is an AI agency in the making. Each bridge is one sector, and Bridge 01 is India's construction and architecture sector. BuildBridge Studio is its product. One person directing a team of AI agents built it. I set the direction, made the engineering calls and own the failures. The agents, running on five Claude models, wrote most of the code; separate agents reviewed it and argued the architecture.
The full write-up will be published on GitHub.
All research notes