Half a million drawings, one model the architects can argue with.
Atelier Foundry is the work file for an eight-month commission with a mid-sized London architectural atelier whose archive runs from 1971 to the present. The atelier had — like several practices we have spoken to in the last two years — a thirty-five-year archive that was effectively inaccessible to anyone who had not personally drawn the buildings in it. The work file is about turning the archive into a working colleague the partners can interrogate during a design week.
An archive that had stopped being read.
The atelier — we are not naming them, at their request, until they choose to do so themselves — came to the studio with a problem familiar to most practices of their generation. Their archive contained roughly 530 000 individual drawings, model files, structural diagrams, and survey notes accumulated across fifty-four years of practice. The founding partner had retired in 2018 and the institutional memory of how the archive was organised had retired with him. Even drawings from projects completed in 2014 were now hard to find without phoning a former associate.
The brief came over two long surveys at the atelier's office in Clerkenwell: build us a way to walk back through our own work. Make it so an architect on a Monday morning, looking for "that thing we did with a brick soffit in a courtyard in 2008", can find it. Eira Ashlin took the file as corridor lead; Pita Selafyan handled the data licensing (this corpus contained sensitive client information and the licensing took the longest of any of our files to date); Wren Halberd helped with the typographic-and-annotation conventions because the atelier's drawing style was itself part of the searchable signal.
What the archive actually contained.
The first three months of the file were not generative work at all. They were digitisation, scanning, structural OCR (for the title-block metadata on the older sheets), and the very slow process of building a metadata schema that the practice's surviving partners would actually agree on. We did not subcontract any of this work; the partner on the file ran it.
Why the work file is called "Foundry".
The atelier's first instinct was to call the deliverable an archive search tool — effectively, a museum index. We pushed back, gently, over two surveys. A museum index makes the past inert; you can look at it but you cannot work with it. The atelier did not need a museum; they needed a foundry — a place where past work is taken down, melted into questions, and re-poured into the work currently on the bench.
The naming choice mattered because it changed what we built. A museum index would have been a search engine over the digitised archive plus a thumbnail viewer. A foundry is a generative model that has read the entire archive and that the architects can argue with: show me this scheme but with the courtyard rotated, show me the brick soffit detail but adapted for a steel structure, show me the language of our 1990s housing schemes applied to this site. The model is not the archive; it has read the archive.
A 780M-parameter spatial-generative model on a single workstation.
The shipped model is a 780M-parameter multi-modal model that reads architectural plans (vector and raster), elevations, sections, and structural diagrams in a shared latent representation. It was trained on the digitised archive over six weeks on a four-A100 box that we rented from a local-to-Bristol university partner with a written training-data licence. The trained model itself is small enough to run inference on a single A6000 workstation, which is what the atelier owns.
The thing the model does well, and that we are most pleased with, is the practice's house language. The atelier has a distinctive way of resolving the meeting of brick, timber, and metal — a recurring vocabulary of soffits, reveals, and steel-supported timber beams — that recurs across fifty years of work in slightly evolving form. The model has learnt this vocabulary as a coherent dialect; it can generate variants that the partners read as "ours" even when the actual building shown has never existed.
The thing the model does badly, and that we have written down, is structure. The model produces plausible plans whose structural logic is occasionally wrong — cantilevers that would not stand up, lintels that span more than they should. We have been explicit with the atelier and the consulting engineers that the foundry is a design tool, not a structural one. The structural engineer's drawing is the structural drawing.
An architect on Monday, with a question.
The interface is, in deliberate contrast to a chat box, structured around a single survey-room metaphor: the architect places a brief on the table (typed prose, plus optional sketch on the digitiser), pins one or more archive references that should anchor the response, sets the constraints of the new site (boundary, levels, programme, planning context), and asks the foundry for six variant studies. The studies arrive as drawings in the atelier's house style, with title-blocks signed "foundry / [partner] / [date]".
The architect can then argue with the foundry — this is too symmetric, ours is more asymmetric than that, the soffit detail is wrong, look at IV.c-2014-038 — and the foundry will iterate. Every iteration is logged. Every signed variant is committed to a new section of the archive marked "foundry-generated"; the atelier's policy is that no foundry-generated drawing is ever issued for construction without a human architect having re-drawn it.
Eight months. Six partners. One quiet rollout.
The foundry shipped in February 2026. The first month of operation was a closed pilot with two partners and one senior architect, deliberately small. The second month opened to all six partners; the third month opened to the senior architects. We are still inside the planned phase I, and the atelier has not yet decided how widely to roll out access internally. The studio's position is that this decision is theirs alone; we do not push on adoption.
The kill switch on this file is the most carefully built of any work we have done to date. The atelier's client confidentiality obligations are serious; if a future regulatory regime, or a client request, required the model and its training data to be removed, the rewind has to be complete and verifiable. The kill switch (foundry rewind --witnessed) requires two signatures, produces a cryptographic attestation of removal, and has been rehearsed three times under audit. We do not expect it to be invoked. We built it as if we did.
Phase II, currently in survey, will fold in the consulting engineers' archive on a separate licensing track. Whether it happens, and on what timeline, is a question for autumn 2026.
If your archive has stopped being read.
Atelier Foundry is the corridor for practices — architectural, civic, museum, urban-planning — with deep archives that have become inaccessible by virtue of size, age, and the retirement of the institutional memory that organised them. We do not take this corridor on for clients without an archive, or for clients whose archive is shorter than ten years; the work depends on the existence of a corpus that has accumulated a real house dialect.