$ cat projects/haifa-hivemind/README.md

Haifa HiveMind

2026ai

A private, local AI research assistant — drop in your papers, ask grounded questions with citations, and it learns from your feedback. Runs entirely offline on your own machine.

Open the live site View the source

Haifa HiveMind — the honeycomb mark on the Odyssus theme

$ less about.txt

Built for a microbiology researcher who needed an assistant that never sends data to the cloud. HiveMind ingests PDFs, DOCX, text and images into a local knowledge base, answers questions grounded in those sources with citations, and reads screenshots with a vision model. When an answer misses, the user rejects it and the assistant rethinks; approved answers feed a QLoRA fine-tuning loop, so it adapts to how the researcher works. Chat history is searchable with a 30-day recycle bin, and simple On / Pause / Off power controls keep the GPU busy only when needed. Shipped as a cross-platform Electron desktop app over a FastAPI + React core — API-first, so mobile clients can connect over the LAN later.

Haifa HiveMind architecture — Electron/Mobile clients over a local FastAPI backend, ChromaDB + SQLite, and an Ollama model server, all on-device

$ less about.txt

$ open architecture.webp