CasaAI | Devpost

Inspiration

The number of home sales increased by 113% year over year in India what’s more, the industry is short about half a million architects, according to India’s Council of Architecture. This results in the urgent need for reliable, quick and environmentally conscious construction. Also on the other hand the average person finds it difficult to visualize their dream home without enlisting the help of an architect which can be an expensive and time consuming affair. Furthermore there is no effective way to view these architectural drawings in 3D for easy visualization.

What it does

By combining user preferences with AI-generated architectural expertise, CasaAI aims to create personalized, eco-friendly, and efficient housing designs that cater to individual needs while promoting environmental responsibility. It primarily allows the average person to create a design that aligns with their tastes. Furthermore these designs can be shared and collaborated on. Architects can speed up design process by generating sample plans for any type of constraints. Commercial buildings can improve sales and customer interaction by using AI’s learning capabilities to enhance floor plans. It will allow for maximizing space and reducing resource usage for lower environmental impact. Generated floor plans can then be viewed in 3D or as a AR/VR which allows for better visualization.

How we built it

CasaAI will use Stable Diffusion an open source image generation model as its base. It will be trained on by manually creating datasets of images of architectural plans using LoRA training methodology. User will specify requirements in a front end interface which will then perform an API call to the model for image generation. Several images will be displayed and user can pick suitable image. For user interaction, CasaAI will strategically employs Next.js and Flask/Express.js for linking the front-end and back-end of the application. The entire processing will be ideally done in the cloud with user preferences and historic blueprint stored in a MongoDB database for easy access.

Challenges we ran into

Image generation models are not as mature as their LLM counterparts and have issues causing generation of significant amount of noise. Models are also prone to ignoring user specifications which requires more training and higher quality datasets. Defining accurate and comprehensive sustainability metrics to assess design efficiency and environmental impact is difficult.

What's next for CasaAI

Currently AI model is being run locally which results in slow load times. To improve this the entire model and interface will be hosted in the cloud.
Continuous Improvement of AI Models: Regularly update and fine-tune the AI models using new data and user feedback. Incorporate advanced techniques like transfer learning and neural architecture search to improve the quality and diversity of generated designs.
Virtual Reality Integration: Enhance the virtual tour feature by integrating more advanced virtual reality technologies, allowing users to experience their designed homes in even greater detail and realism.
Data-Driven Insights: Aggregate and analyze user data to gain insights into design trends, user preferences, and sustainable housing demands. Use this information to inform future updates and feature enhancements.
Currently only trained with residential housing plan data which proves ineffective in designing commercial establishments. So improvement to model for commercial applications is necessary.