Independent AI Interaction Prototype 2026

DesignAlignmentAgent

Can AI explore visual preferences through selection, reaction and feedback instead of relying on prompting alone?

An independent AI interaction prototype that turns visual choices and natural-language feedback into refined design directions and a structured final brief. Interior design is the first test case, but the same approach could be adapted to brand identity, poster design and other visual design tasks.

See the interaction
01 /

How the interaction works

The working prototype, from first selection to final brief
Selection → generation → natural-language response → refinement → brief
Additional visual evidence

Exploring directions

A sequence of interior-design variations from the prototype, available to play on demand. This shows the visual range; the screen recording above shows how the user interacts with the system.

02 /

The question behind the prototype

Problem

Visual taste is subjective. People can recognise what they like but may struggle to describe a complete direction from a blank page.

Hypothesis

Comparison and natural-language reaction may reveal useful preference signals that a single upfront prompt cannot capture.

Prototype

A two-round interaction tests whether those signals can guide increasingly focused visual options and a usable final brief.

03 /

User journey

From user action to application behaviour to model contribution

This diagram shows what the user does, what the application handles and what AI contributes. Each round builds on the references and feedback already collected.

00   Reference foundation10 handcrafted visual styles, available before the session begins
Minimalist
Maximalist
Industrial
Scandi
Japandi
Midcentury Modern
Modern Farmhouse
French Country
Hollywood Regency
Coastal
01

Selects two styles

Chooses a primary and secondary reference from the foundation.

Stores the selection

Creates a style pair and starts the first generation round.

Interprets and plans

Combines the references and plans four new directions.

02

Reviews six directions

Two selected references plus four new generated options.

Renders progressively

Requests the images sequentially and displays each as it arrives.

Generates four images

Renders additional options grounded in the style pair.

03

Describes the response

Explains liked and disliked elements in natural language.

Preserves the context

Associates feedback with the directions shown in Round 1.

Extracts signals

Interprets preferences, dislikes and overall mood for refinement.

04

Reviews three refinements

Compares a smaller set of directions informed by the feedback.

Starts Round 2

Passes the prior context and renders the refined set sequentially.

Plans and renders

Converts signals into three refined visual alternatives.

05

Selects the closest result

Chooses the preferred Round 2 option.

Triggers the final brief

Retains the chosen option and session history for synthesis.

Synthesises the outcome

Produces a title, summary, design rules and preference scores.

Preference emerges through a sequence of decisions, not a single isolated prompt. The system keeps prior selections and feedback available throughout the session.

04 /

The interaction is the product

Interior design is where I am testing the prototype today. By changing the starting references and design prompts, the same selection-and-refinement process could be adapted to brand identities, posters and other visual design projects. These are potential applications, not versions already built.

01

Pick two references

02

See new directions

03

Describe what works

04

Review refinements

05

Receive a final brief

Current domain and possible applications

Built test domain

Interior design

Potential application

Brand identity

Potential application

Poster design

Potential application

Visual identity systems

05 /

System architecture

Documented V2 model-call structure, simplified for reading

Lightweight interpretation and planning stages are separated from image generation. Feedback shapes the second render round, and session state persists throughout. The documented V2 sequence uses five text calls and seven image calls. The prototype uses Gemini, Python/FastAPI, Firestore and Google Cloud Run.

Call 01 · Text

Interpret styles

Read the primary and secondary starting references.

Call 02 · Text

Plan directions

Prepare four Round 1 image specifications.

Calls 03–06 · Image

Render Round 1

Generate four directions in sequence.

Call 07 · Text

Analyse feedback

Translate reactions into structured signals.

Call 08 · Text

Plan refinements

Use the signals and first-round references.

Calls 09–11 · Image

Render Round 2

Generate three refined directions in sequence.

Call 12 · Text

Synthesise brief

Summarise the selected direction and preferences.

Output

Structured brief

Title, summary, design rules and signal scores.

Persistent state · Google Cloud FirestoreSelections, feedback and generated results remain associated with the session.
Designing around real constraints

Image-generation requests were sequenced with cooldowns to respect API rate limits. This introduced an interaction trade-off: generation took longer, but cards could appear progressively instead of requiring the user to wait for the complete set. These timings describe the documented V2 implementation, not a benchmark of a production service.

06 /

What I designed and built

Interaction model

Designed the reference-selection, natural-language response, refinement and final-choice sequence.

System behaviour

Defined how interface actions, saved context and model calls progress through a multi-round session.

Reasoning structure

Separated interpretation, planning, feedback analysis and final synthesis from the image-rendering stages.

Prototype implementation

Built and deployed the V2 prototype using AI-assisted development, Gemini, Python/FastAPI, Firestore-backed session state and Google Cloud Run.

Visual and UX direction

Designed the interface and system documentation around a visual decision-making problem.

07 /

What I would test next

The current prototype tests the interaction in interior design. The next questions are whether this process produces useful alignment more reliably than a blank-prompt approach, whether one round can work as well as two, and how the same mechanism transfers to another visual domain.

Development reflections

Notes from building the prototype, published on LinkedIn.

From Renders to Reasoning ↗

Building AI-Powered Systems: Lessons from Practice ↗

Building V2: What I Learned About Limits, Patience, and Platform Choice ↗