AI & Tools
Local AI Chat
A private ChatGPT-style workspace that runs LLMs entirely on local hardware — no data leaves the machine.
About the project
A chat web app for teams that need AI but can't send company documents to cloud services. The Next.js app streams responses from Lemonade Server, an OpenAI-compatible model runner using llama.cpp on an NVIDIA GPU. Users choose Lite, Flash or Pro tiers backed by Qwen models of different sizes. There's more to it than chat. A knowledge base does retrieval-augmented generation (RAG) over uploaded PDF, Word and Excel files using a local Thai-capable embedding model. A Python sandbox (Pyodide on WebAssembly, with numpy and pandas) lets the model calculate and analyse data safely. Voice mode uses Typhoon Whisper for Thai speech-to-text and a Thai TTS server. There's also meeting transcription and a Concept Studio for generating architectural concept images. An admin console manages models and installs them in one click. Benchmarking Thai speech models to fit within 8 GB of VRAM alongside the LLM shaped many of the model choices.