Celeste
A fully local AI desktop assistant — shipped as a real installer, not a dev script.
A privacy-first desktop AI assistant for Windows and Linux. Inference, memory, and document search all run on the user's own hardware after setup — no cloud calls, no subscription.
// Architecture
llama.cpp via three swappable backends (in-process, subprocess, HuggingFace transformers), configurable GPU offload
Two-tiered — episodic conversation history in a local Chroma vector DB, plus a SQLite subject-predicate-object fact graph
Hybrid document Q&A — an always-on TF-IDF lexical index plus an optional multi-GPU semantic deep index
PyInstaller + Inno Setup for Windows, tarball/AppImage for Linux — a real installer with a first-run GPU-detection wizard
// What's Actually Hard Here
Automatic GPU vendor detection (NVIDIA, AMD, or CPU fallback — Vulkan vs. ROCm) baked into first-run setup is easy to skip in a dev script and genuinely hard to get right across unknown consumer hardware in a shippable installer. This is the proof point for shipping real installable software, not just a working demo — it's sold as a real product.