DesinIt.AI
- pytorch
- python
- langchain
- fastapi
- postgresql
- prisma
- nextjs
- typescript
- tailwindcss
Introduction
Architects spend a huge chunk of the early design phase manually drafting and iterating on candidate floor plans — slow, repetitive, and a real bottleneck on large construction timelines. Worse, existing generative floor plan tools are trained on Western datasets and don't reflect the spatial conventions of Algerian residential architecture.
DesinIt.AI closes that gap. Describe an apartment in plain language — "a two-bedroom with an open kitchen facing the living room" — and get back a real, CAD-compatible vector floor plan, no drafting required. Built as my End of Studies engineering thesis, in partnership with the Hasnaoui Group (GSH), one of Algeria's largest construction and real estate groups.
🌟 Key Features
🧠 Natural Language to Floor Plan
- Conversational Generation – describe an apartment in free-form text, get a structured room-adjacency graph via a fine-tuned LLM.
- Iterative Refinement – a ReAct agent lets you adjust the generated layout through follow-up instructions instead of starting over.
- CAD Export – every generated plan exports directly to DXF, ready for real architectural workflows.
📐 Custom Diffusion Model — Plan-Diffusion
- Boundary Conditioning – constrains the generated layout to a user-specified building footprint.
- Per-Room Area Conditioning – specify target room sizes directly instead of hoping the model gets it right.
- Domain-Adapted – fine-tuned on a proprietary archive of real Algerian floor plans provided by GSH, on top of the public ResPlan dataset.
🛠 Tech Stack
- ML & Training: PyTorch, Unsloth (LoRA fine-tuning), vLLM (multi-adapter inference), Weights & Biases, LangSmith
- Agent & Orchestration: LangChain (ReAct agent)
- Backend: FastAPI, PostgreSQL, Prisma
- Frontend: Next.js, TypeScript, Tailwind CSS
- Data Pipeline: AutoCAD, Photoshop, OpenCV, NumPy
📊 Results
- Fine-tuned LLM candidates reached near-perfect schema validity and room-type accuracy above 0.98.
- Plan-Diffusion reached graph compatibility within ~17% of the original HouseDiffusion baseline — while adding boundary and area conditioning it didn't have.
- Evaluated zero-shot on Algerian floor plans (no local fine-tuning at all), the system still generated a clear majority of typologies as coherent, valid apartments.
🎓 Academic Context
Defended June 2026 at École Nationale Polytechnique d'Oran — Maurice Audin, for the State Engineer's Degree in Information Systems. Built with my thesis partner, Abdessamed Benaidja, under the supervision of ENPO-MA and GSH.
Repo is currently private due to the industrial partnership — happy to walk through the thesis or a live demo directly.
Built with 🧠 for architects who'd rather describe a floor plan than draft one.