Rose's Corner

A restaurant above a Redfern corner, designed with AI at every step.

Exploded isometric of Rose's Corner: the parapet lifts to show the room above the street

Sohan Mitra · AI for Architects 2026 · For practice partners and directors

What I learned on the AI course, and how the practice can use it.

One project, Rose's Corner, taken through five assignments. Each one taught a way of working that can carry over to real jobs.

How it can work in practice

Before the first meeting

Run any new address through the site app.

During design

Send white cards from Rhino or Revit straight to rendering.

On a live job

Drop saved emails and the programme into the coordination app.

Always

The architect checks, and the verdict stays in one tracker.

Why this matters

AI takes the repetitive work. The architect keeps the decisions.

01 · Listen

Style summary

02 · Imagine

Concept

03 · Read

Site reading

04 · Draw

Design resolution

05 · Coordinate

Coordination

Input
An evening's interview
The brief, the site and a hand-drawn base
Open data and site photos
Concept and style rules
Consultant emails
AI does
Themes, style research, images
Concept images over the hand-drawn base
A live map with noise built in
Plans and sections through Revit and Rhino
Quotes, flags and raises RFIs
I check and seal
The ethical rules
The concept: the window as portal
What to check on site
Every design decision
Every RFI
Every output comes back to the architect to check and seal
01

Style summary

One interview, turned into themes, then a style.

Wispr Flow, Claude, Gemini

What it showed me: Kolkata can come in as real materials, like a rammed-earth bar front, not as pictures of Kolkata.

  1. Wispr Flow

    Transcribed the interview live, over one evening.

  2. Claude

    Drew out themes, not solutions, then the style research.

  3. I check and seal

    Caught the "village look". Wrote the rule: nothing depicts rural poverty.

  4. Gemini

    Made the images, from prompts written to the themes.

  5. I check and seal

    A named source for every claim, then back to Arshia.

02

Concept

The window as portal.

Claude, Higgsfield

What it showed me: the windows can do the work of a shopfront. The street glimpses the room but can't walk in.

03

Site reading

A live map the team can use, with the noise reading built in.

CoWork, archiMap, Site Analysis Studio · NSW planning, open data

What it showed me: one fixed map lets several site issues be discussed at once, with each photo where it was taken.

Reflection: a live site analysis. What I learnt by building the site analysis as an interactive map dashboard. The live site analysis dashboard
The live site analysis dashboard: map of the corner with layers, and the story panel
Reflection, then the dashboard, the standalone app and my source checks
04

Design resolution

Sheet BD-A1, vector. Pick a drawing and it zooms in; drag to pan, scroll to zoom.

Revit, Rhino, Rhino.Inside, Higgsfield

What it showed me: the backlit stained-glass frieze can be built. It hangs from the structure, is lit from a brass channel and glows on both faces.

Drawing sheet BD-A1: bar section and backlit stained-glass detail, six drawings
Rendered views
How it was achievedI sketch or brief, Claude builds and checks its own work, I redirect or reject.

  1. ModelRounds 1–19

    Site and room in Rhino, pushed into Revit through Rhino.Inside.

  2. My sketchesRound 22

    Bar cabinet, curved counter, glass frieze, coved ceiling, banquettes.

  3. Build and detailRounds 22–26

    Native Revit families, then sheet BD-A1, laid out to my markup.

  4. White card and model views

    Exported straight from Rhino or Revit.

  5. Claude Higgsfield connector

    White cards go in. Prompts carry the materials.

  6. Rendered views

    I check each one against the model, and reject what drifts.

05

Coordination

The AI quotes and flags, and I decide.

Claude, Excel, coordination app · Email

What it showed me: every decision can be traced, from the RFI raised to the answer, with the consultant's words kept exact.

  1. Consultant emails

    Eight consultant agents email and raise RFIs: 75 so far.

  2. AI reads every email

    Into the Excel tracker: words copied exactly, inferences in a separate column.

  3. I check and seal

    Each RFI in the tracker: accept, accept with action, amend, or reject with reason.

  4. Dashboard pulls it through

    Who owes answers, what is open (57 of 75), what is due next.

  5. Trace and copy

    Open any decision: when, where and who. One click copies it for an email.

The coordination dashboard: open questions, progress, who owes answers, decisions and next dates
All emails and the programme are hypothetical, written for this course.

Practice partners and directors, my employer. They decide if this fits the office.

Awards submission

  • Leads with: a completed project and the registered architect behind it.
  • Shows: selected images, drawings, plans and a sustainability checklist.
  • Leaves out: unbuilt work.

Practice AI adoption

  • Leads with: policy and a roadmap with go/no-go gates (Flad, HOK).
  • Shows: repeatable capability and office-wide critique (Olson Kundig).
  • Leaves out: one-off experiments.

Client pitch

  • Leads with: the client's situation.
  • Shows: relevant proof, then the ask.
  • Leaves out: firm history and logos.

What I take: lead with the outcome, show proof and who checked it, leave out tool tours.

Sources: AIA NSW awards entry · aec+tech on Flad, NBBJ, Stantec · Dezeen on HOK · Olson Kundig · pitch guide (general, not architecture-specific). Awards image counts and word limits are not verified.

Same components, three audiences.

Other arrangements, brainstormed with Claude.

Practice partners (chosen)

Why this matters, then the five components in workflow order, then my reflection.

Kept because partners judge whether it fits the office, so the outcome comes first and the cost of checking stays visible.

The client, Arshia

Her memory, the room, the interior renders, then what happens next. Coordination and site data move to a back page.

Awards jury or editor

One hero image and the concept first, then the design resolution. Process becomes a short appendix.

Claude built this deck from my folders, to a brief I set and corrected.

What Claude was told

  • Audience: practice partners and directors.
  • One idea per slide, minimal scrolling. Small text; the work speaks louder.
  • Order: the five components in workflow order, then the reflection.
  • Style card: sand paper, terracotta for lines only, no blue.
  • Name every gap. Don't invent figures.

Rounds

Sixteen published versions in one Claude Code session. Each round was my review of the page, then a list of changes.

Guardrails that carry over

  • Ask for themes, not solutions.
  • Write the "no village look" rule before the image prompts.
  • Agents copy consultants' words exactly and keep inferences in a separate column.

Reusable on another job

The eight-agent, 75-RFI coordination setup can run again: swap the brief, the drawings and the consultant list.

A short AI policy for the practice: three go/no-go gates.

Each gate is a point where I stopped the AI on this project. Each one can be adopted as it stands.

Style rule

A person enforces the brief's style rules, such as "no village look".

Go
The output follows the rule.
No-go
Output that breaks the rule is rejected.

Planning status

Anything flagged as possibly lapsed, like the 2021 consent (HER-01), goes to a person.

Go
A person has checked the status.
No-go
The AI never updates a status silently.

Third-party data

Heights taken from OpenStreetMap stay marked indicative.

Go
Checked against a council record.
No-go
Treated as fact before that check.

What the AI did, what I did, where the seal sits. The AI ran the testing from my inputs. I took the time to observe, debug, suggest and make changes. My seal sits at the critical items, where a design decision matters most: the three gates above. I checked what went out as well as time allowed, and some items were missed or not seen properly, as the assignment remarks reflect. AI processes a lot of information and hands a lot of it back for the human to check and seal. That is where the time goes.

Prepared by Sohan Mitra · AI for Architects 2026 · The client and consultants are fictional personas

Reflection: where the time goes.

My final reflection, then six moments from the working chat that show where my judgement went. Turn the page, or click a spread to read it full size.