sebkuepers/Ongiini

A free and open AI Assistant for Namibia # Ongiini > "Ongiini" — "How are you?" in Oshiwambo. A free WhatsApp AI assistant for people in Namibia. English & Afrikaans today, Oshiwambo via a translation layer coming soon. Powered by **Gemma 4 26B A4B** (NVFP4) running locally on an NVIDIA DGX Spark, exposed via a Cloudflare Tunnel. Pilot currently runs in Germany; the goal is to move the hardware to Namibia once the service is sustainable. Ongiini is the first project of the **Common Intelligence Foundation** — a non-profit foundation currently being formally established in Estonia. Until that registration is complete, operations are carried out by Sebastian Küpers in his personal capacity, on a non-profit basis. See common-intelligence.org for the manifest. The foundation's website source lives under `foundation/` in this repo but is **gitignored** (kept local-only until the foundation is formally registered). It is deployed to Cloudflare Pages directly via `wrangler pages deploy`. ## Built on Owela The chat-agent framework that powers Ongiini lives in this repo as a sibling package: `owela/`. Owela is an opinionated framework for small open-source models (Gemma, Llama, Qwen) on inference engines (vLLM, llama.cpp, ollama) delivered through messenger transports (WhatsApp, Signal, Telegram) — the deliberate opposite of frontier-model-via-API frameworks like LangGraph. It was extracted from this codebase as the patterns repeated themselves and the application file grew past what one file should hold. Owela is MIT-licensed and is kept embedded in this repo for v0; if a second consumer emerges it can split to its own repo + PyPI. See `owela/README.md` for the framework's positioning and `ongiini/README.md` for the technical counterpart to this operator manual. ## Architecture ``` WhatsApp Cloud API ──▶ api.ongiini.ai │ (Cloudflare Tunnel) ▼ 127.0.0.1:8445 ──▶ FastAPI webhook (Docker) │ ├─▶ vLLM Gemma 4 (host:8124) │ ↳ text + image (multimodal) ├─▶ Tavily (web search) ├─▶ mem0 long-ter …