M2 Master Agentic AI in virtual worlds An Interactive AI Experience for Sustainable Nutrition "A Better Bite"

Miro
Used for early brainstorming and concept development. Its infinite canvas was ideal for mapping out complex conversational user flows and documenting design decisions simultaneously.

Mermaid
Used to create system architectures and flowcharts directly from text-based syntax. This ensured our technical documentation was version-controllable and easily maintainable as the AI evolved.

GitLab
Our primary platform for version control, code reviews, and issue tracking. It allowed the team to efficiently manage the separate workflows of the Unity frontend and Python backend.

Discord
Our primary communication platform. It provided structured channels and video calls to ensure seamless daily collaboration and resource sharing between frontend and backend developers.

Unity 6
Selected for its advanced rendering pipelines and performance optimizations, enabling a high-quality 3D kitchen environment that remains smooth while handling background AI processes.

C#
The mandatory Unity scripting language. We heavily utilized its asynchronous features (async/await) to orchestrate non-blocking REST API calls, ensuring UI fluidity during real-time LLM communication.

FastAPI
A high-performance Python web framework used to build the REST API connecting Unity to the AI services, chosen for its native asynchronous support.

Uvicorn
A lightning-fast ASGI web server used to run FastAPI, ensuring the high-performance request handling required for real-time game interactions.

Python
The industry-standard language for AI, providing the extensive ecosystem necessary to seamlessly integrate our LLMs, LangChain, and audio processing libraries.

ChromaDB
An open-source vector database powering our RAG pipeline. It stores embeddings for recipes, kitchen details, and F4F research, allowing the LLM to retrieve verified context and minimize hallucinations.

Groq Cloud
Provides specialized hardware acceleration for AI inference. We used it to access Llama 3.1 8B Instant, minimizing latency for responsive, real-time voice dialogue.

LangChain
A framework used to orchestrate prompts, integrate our retrieval database, and manage session memory so the AI remembers project context and recipe steps.

Faster-Whisper
An optimized reimplementation of OpenAI’s Whisper, enabling fast, accurate, local transcription of player voice input for hands-free conversations.

Piper
A fast, local text-to-speech engine. By utilizing its diverse predefined voices, we enabled dynamic character randomization in Unity while avoiding cloud API latency and costs.

Hugging Face
A central hub for machine learning used to source and deploy the specialized pre-trained models required for our audio processing pipeline.