🜂

acknowledgements

This whole surface stands on open source. Named, with thanks — because that's the honest thing.

Infrastructure

Caddy — automatic HTTPS and a reverse proxy that just works; it fronts every request here.

Let's Encrypt — free TLS for everyone. The reason https on a bare box is a given, not a luxury.

Redis — the instant 65-second cache in front of the free tier; strong read-after-write where a global KV couldn't.

systemd — keeps the gateway, model servers, and dogfeed loop alive and restarting.

Debian / Ubuntu & Linux — the ground everything runs on.

Serving & models

Ollama — dead-simple local model serving; the free tier's Qwen3-Coder-30B runs on it.

vLLM — fast, batched, OpenAI-compatible GPU serving; the paid H200 runs on it.

Qwen3-Coder (Alibaba, Apache-2.0) — open, dedicated coder models you can read the weights of. The heart of every tier.

Hugging Face — hosts the models and the open dogfeed corpus; the Hub API carries every published batch.

Runtime & tooling

Python — the gateway, the RIVA server, and most of the glue.

Bun — the fast TS runtime the dogfeed loop runs on.

SQLite — the zero-ops store the dogfeed loop records into.

uv (Astral) — fast, sane Python packaging — the RIVA venv is built with it.

Frontend & assets

Pillow — generated the OG image, the product cards, and the favicon.

ImageMagick — packed the multi-resolution favicon.ico.

Vanilla HTML/CSS/Canvas — no framework, no CDN, no tracker. The neural-field background is ~60 lines of hand-written canvas.

RSS — in memory of Aaron Swartz (1986–2013), who co-wrote it and believed information is for everyone. He was right. This open corpus is for that idea. 🜂