PhysiCast est un moteur d'IA physique développé pour démontrer des capacités avancées en perception visuelle et en raisonnement spatial.

PhysiCast is a Physical AI prototype designed for real time scene interpretation and spatial reasoning. By bridging computer vision and localized large language models, the system converts raw environmental data into actionable Spatial Knowledge Graphs.

• Scene Interpretation: Leverages YOLOv8 to extract precise 2D object coordinates and semantic labels.
• Spatial Reasoning: Utilizes Llama 3.2 (via Ollama) to infer physical relationships, proximity, and interaction zones.
• Edge AI Architecture: Built with FastAPI to ensure low latency and data privacy through local execution.

1. Input: User uploads a physical scene image via the web dashboard.
2. Vision Engine: YOLOv8 detects objects and generates spatial bounding box data.
3. Reasoning Engine: A customized prompt feeds coordinates into a local LLM for context aware interpretation.
4. Output: A semantic description of the physical space for robotic or digital interaction.

• Language: Python 3.12
• Framework: FastAPI, Uvicorn
• Computer Vision: Ultralytics YOLOv8
• Inference Engine: Ollama (Llama 3.2)