Cognilumin

A few more behind‑the‑scenes details:

The $0 operational cost is real — but it required writing a PHP proxy that rotates between three different AI providers. If one returns 429, the next tries after a 1‑second delay. If all three are exhausted, it falls back to OpenRouter's free Llama model. This chain took about 3 hours to get right and has never failed in production.

The Illuminator's system prompt is ~200 words and strictly enforces a pedagogical structure: mental models, analogies, key takeaway. Getting the LLM to follow that format across providers was harder than I expected — Gemini follows it perfectly, Groq sometimes ignores the "analogies" part, and OpenRouter's free model needs a nudge.

I built this on InfinityFree shared hosting, which means there's a cookie‑based anti‑bot challenge on every request. I had to route traffic through Cloudflare to make the API endpoints reachable. If you're building on free hosting, that's one to watch.

Happy to answer any technical questions!