Work · What it looks like
What the work looks like, in pictures
Short answer: Six kinds of job our agents do, each with a few images of the setting and the output. The images are generated for this page and labelled; they are not photographs of a customer.
Three of the six are things we have built and run (labelled Fact); three describe the scope of work as it has been put to us, before a customer has measured it (labelled Illustrative). No customer, supplier or private person is shown or named.
Updated 3 November 2026 · images made on our own GPU server with open-weight models (Z-Image, FLUX.2 klein, Qwen-Image, Wan 2.2) and chosen by a vision model plus a person
A research database of suburban retrofits
Fact A database of more than 2,500 suburban retrofits: dead malls, office parks and strip centres, and what became of them. Extraction workers read the sources, an API serves the records, and a public site lets a researcher’s archive be searched by place and type. Built for a university partner; the write-up appears when the partner says so.
A family photo library, kept in the house
Fact A household’s 33,844 photographs, ranked by an aesthetic model, grouped into albums and restored where the originals had faded. Everything ran on a Mac and a workstation in the house; no photograph left the building.
A quote desk for a cabinet shop
Illustrative A quote desk and inbox triage built from the shop’s own price list. Each request becomes a priced draft with the price-list row shown beside every line, so a person checks the numbers and sends the quote. No customer has run it yet; the measured results on fictional companies are on the scorecard.
Deal analysis and diligence for a family office
Illustrative Lease abstracts, rent-roll questions and loan-document checks answered from the office’s own files. The output is a diligence register and memo drafts for the investment committee, with every answer pointing back to the page it came from. Scope as described to us; not yet measured on a customer’s files.
A public-company profile library for a deal team
Illustrative 124 public-company profiles of about 10,000 words each, assembled into one searchable brain. The deal team asks questions in plain language and gets answers with the profile and passage they rest on. Scope as described to us; not yet measured on a customer’s files.
An on-premises unit, and making it fast
Fact A Mac mini or a GPU workstation in the customer’s building, running their own models on their own data. Tuning our own server took one model from 48 to 173 tokens a second without new hardware (30 October 2026); the method is on the model-optimisation page.
How these pictures were made, and what they are not
- Every image was generated on our own GPU server from a written brief, in a documentary style, with no text, logos or faces in close-up. Four open-weight image models and one video model produced about 2,600 candidates; a vision model scored them and a person picked.
- They show the kind of place and the kind of output. They do not show a customer, a customer’s files, or a real result. Numbers on this page carry their own label and source; none comes from an image.
- Where a write-up exists with dates and sources, it is on the Work page; this page is the picture version.
Or call us: 404 493 6248 · Email sales