The AFlow — Operator

01 / THE PROBLEM

Three gaps in AI-assisted design

Architects use AI tools every day. The tools forget. They don't track who made what. They can't preserve what matters.

01 / MEMORY

Every AI call starts from scratch.

You generate a concept, then re-describe everything — the massing, the site, the materials — because the next tool has no idea what came before.

02 / AUTHORSHIP

When AI generates, who made it?

What was the input? What was the reasoning? In education and practice, design provenance matters — for assessment, for IP, for understanding the design process.

03 / GOVERNANCE

Everything is fungible by default.

You can't say "keep the massing but change the facade." There is no way to lock identity-defining features while exploring others.

02 / WHY AFLOW

Not a chat box. Not a generator.

Why use The AFlow instead of a GPT or an image AI tool? One great image isn't a design — the part that matters is everything those tools throw away the moment you start the next prompt.

ChatGPT and image tools are excellent at producing one output at a time. The AFlow is a design-reasoning environment: generation is one component inside a process that remembers your intent, records who made what, and lets you lock the parts that define a design while you keep exploring the rest.

What designing actually needs

The AFlow

Chat assistants
(ChatGPT)

Image / design-AI
(xFigura, generators)

Remembers your design intent across iterations

✓

one chat

—

Tracks who & what made each output (authorship & provenance)

✓

—

Lock identity-defining features while you explore the rest

✓

—

Reasons over geometry & spatial structure

✓

—

Text, image & 3D in one connected workflow

✓

text only

partial

Built for studios & teaching cohorts

✓

—

03 / HOW IT WORKS

Create · Govern · Trace

A visual node graph where every operation is a design decision — and every decision is preserved.

STEP 01

Create

Connect nodes: text prompts, image generators, 3D models, LLM reasoning. Each connection is a typed data flow. Each node is a recorded operation.

STEP 02

Govern

Mark identity-defining features as non-fungible. The AFlow preserves them across every downstream iteration — automatically.

STEP 03

Trace

Every action is recorded — who, what, when, why, how much it cost. Browse versions, replay any state, export the full provenance.

04 / FEATURES

Studio & practice

The same governed graph — tuned to how you work. Pick your context: a teaching studio, or a practising office.

For the studio — tutors running cohorts and students handing in a process, not just a render. Everything here ships in v0.7.2 today.

⬡ 01 / AI GENERATION

Multi-modal generation

Text-to-image, image-to-image, sketch-to-image, image-to-3D, text-to-3D, LLM reasoning and critique — all in one graph.

● Live FLUX Gemini Claude Tripo

⬡ 02 / DESIGN MEMORY

Context flows automatically

Relevant history, locks and constraints are brought into each generation automatically — so you never re-describe your project to take the next step.

● Live Auto-context Locks-aware No re-prompting

⬡ 03 / FnF GOVERNANCE

Lock what matters

Toggle any node output Fungible or Non-Fungible. Combine with Mutable / Immutable for absolute or AI-proposed-human-approved anchors.

● Live NF · I NF · M F · M

⬡ 04 / PROVENANCE

Every decision recorded

Every action is recorded — author, input, output, cost, duration and the features it affected. Browse it as a visual 5-layer provenance map, or export the whole trail.

● Live Exportable 5-layer map

⬡ 05 / TOPOLOGY

From mesh to room layout

From mesh to room layout: boolean operations, spatial graphs, mesh extraction, IFC / BIM import and generative subdivision — all on the canvas.

● Live Spatial graph IFC GLB 12 ops

⬡ 06 / FOR COHORTS

Built for studio teaching

Roles for students, tutors and admins; per-cohort budgets and caps; invite-based onboarding; and collaborative sessions.

● Live User / Tutor / Admin Groups

For the office — professional architects and building designers carrying a scheme from sketch to a checked plan. Each tool is marked by maturity: ● Live ◐ In build ○ On the roadmap

⬡ 01 / AI GENERATION

Explore options fast

Concepts, images and 3D from text, sketches and reference — generated inside a governed graph that records every step, not a throwaway chat.

● LiveFLUXClaudeTripo

⬡ 02 / GENERATIVE FLOOR PLANS

From a brief to a checked plan

A governed engine turns a brief into a watertight room layout on a Topologic model — doors, adjacencies and areas included — then writes it out as IFC.

◐ In buildTopologicCellComplexIFC4x3

⬡ 03 / ENERGY AGENT

A second opinion on energy

An agent reads the model and returns a steady-state energy picture — fabric, infiltration and solar by orientation — checked against an EnergyPlus reference. It never reports a number it can't trace to a solver.

◐ In buildEnergyPlus-checkedSourced numbers

⬡ 04 / STRUCTURE AGENT

Early structural sanity

A structural agent takes gravity loads down the wall-and-column graph and reports member utilisation — the same observe → act → evaluate loop, on the same model.

○ On the roadmapGravity take-downMember checks

⬡ 05 / WORKS WITH YOUR BIM

IFC in, IFC out

Bring in IFC, Rhino, Revit and mesh models as topology with metadata. A live 4D building model — timeline, edits and server-checked geometry — is in build.

● Live importIFC4D model · soon

⬡ 06 / ACCOUNTABLE BY DEFAULT

A signed record, locks that hold

Every action is signed into an append-only record — an accountable “golden thread” of design information. Lock the parts that define the scheme and the executor holds them, even when an agent is doing the work.

● LiveGolden threadExecutor-enforced

In build — across the platform ◐ ON THE WAY

◐ Design context preview

See the context informing each AI generation, so you know exactly what the model is working from.

◐ Critique → regenerate loops

Guided iterate cycles on the canvas: review a result, refine the brief, regenerate — without losing the thread.

◐ Stronger feature preservation

Keep identity-defining elements intact across iterations, even as you push everything else further.

◐ Find similar past designs

Surface related earlier work across your sessions, so good moves don't get lost between projects.

05 / DESIGN RIVERS

OntoCAAD concern strands

Six colour-coded concerns flow through your graph. Where they merge, you can see exactly which design questions a node is addressing.

Structure

Load paths, grid, frame

Form

Geometry, massing

Environment

Climate, daylight, plant

Programme

Use, flow, adjacency

Materiality

Surface, texture, finish

Context

Site, streetscape, neighbour

06 / LIVE CANVAS

A node graph you can drive

Toggle the FnF lock and watch downstream nodes update their behaviour. Flow is enforced at the executor.

A lock isn't a note in a file the next step is trusted to honour. When the graph runs, the executor itself refuses any operation that would change a feature you've fixed — so governance holds even when an AI agent is doing the work.

TEXT INPUT

Hospital · stacked podium

TEXT TO IMAGE

Nano Banana

1.5 ✦

IMAGE TO IMAGE

Watercolor style

2.8 ✦

NF · LOCKED

LLM REASON

Claude Sonnet · critique

1.5 ✦

IMAGE TO 3D

Tripo v3 · 21k faces

22.6 ✦

CLICK NF · LOCKED TO TOGGLE

07 / PROVENANCE

Lineage · authorship · cost

Every action is signed. Human or AI agent, what changed, which features were affected, how much it cost.

The same append-only record gives a building an accountable “golden thread” of design information — the kind the UK Building Safety Act 2022 expects to be created, kept and handed on.

Artefact

hospital_section_v2.png

sha-256 · 4a7e…b921

14

Operations

35 ✦

Cost

11

AI calls

3

Author decisions

Features affected

Structure Programme Materiality

Locks

⬢ Massing · NF · Immutable

⬢ Skylight · NF · Mutable

Record

Append-only · signed · exportable

Built for a golden-thread information trail