Rooms App

The problem

Asking a model to furnish a room gives you a nice description and a layout that does not work: pieces overlap, a wardrobe blocks the door, the sofa costs more than the whole budget. A design is only useful if it can be built in the real room, from things you can actually buy.

The solution

You add photos of a room, type its width, length and ceiling height, and describe what you want: the style, a budget, the pieces you must have. A team of model-backed agents and deterministic steps furnishes the room with real catalog items and places them. You get a floor plan, a shopping list with prices and alternatives, the reasons behind the design, a 3D view and the new furniture drawn into your own photo.

• The model proposes, code checks. The Room Analyst, Designer, Architect and chat edit use models. The solver that places the furniture and the Critic that reviews it are plain Python, so a layout is never accepted on the model's word.
• A Critic loop. Overlaps, blocked doors, windows and radiators, pieces nobody can walk up to and the budget are checked every round. Failures go back to the Architect or the Sourcer for up to three rounds, and the best round is kept.
• You confirm the room. The run pauses after the photos are read, so you can fix what the analyst got wrong before any design is made.
• Chat edits are checked the same way. "Make it cheaper" or "move the desk closer to the window" is validated before it is applied, and can be undone.

1. RoomAnalyst, you confirm
   ↓→
2. DesignDesigner, Architect
   ↓→
3. PlaceSourcer, solver
   ↓→
4. CheckCritic, up to 3 rounds

Evaluation

The agent graph is compared with a single-model baseline and with versions of itself with a part removed, on layout metrics and a vision-model judge. A small paid pilot, 12 runs on three rooms, cost about $0.58. Every design stayed in budget, covered every must-have and had nothing out of the room or blocking an opening. The full graph cost about 70% more than the single call and took about three times longer, and the judge could not tell the variants apart. With three rooms and one repeat, these are anecdotes, not results.

Demo