I’ve been building Cognilumin on and off for the past few months, mostly late nights and weekends. It’s a stateless AI observability platform—three tools that help explain, visualize, and debug AI behavior. The catch? It runs on $0/month infrastructure with no frameworks and no paid APIs.
Here’s what it does:
· Illuminator – paste any AI‑related question and get back a structured, master‑teacher explanation with analogies, mental models, and a key takeaway. The backend detects the intent (evaluation, debugging, observability, or general teaching) and selects a specialist prompt.
· Topology Visualizer – describe a system scenario in plain English and a live particle‑based flow‑field engine renders an animated state map. Think lightweight Galileo Smartscape.
· *Execution Logs *– give an AI agent a task and trace every reasoning step with hex‑formatted indices, action tags, and latency profiling. Browser‑based, zero‑setup LLM debugging.
How I kept the stack at $0
· Hosting: shared hosting (moved from InfinityFree to Hostinger for cron jobs)
· AI providers: Gemini, Groq, and OpenRouter—all free tiers, with a PHP proxy that rotates keys and handles 429s automatically
· Storage: none—the platform is completely stateless (no logins, no database, no tracking)
· Dependencies: zero—vanilla PHP + vanilla JavaScript, not a single framework
What I learned along the way
**
· **Rate‑limit juggling is a real skill. Writing the fallback proxy (Gemini → Groq → OpenRouter) took a few hours, and it hasn’t failed once in production.
· Prompt engineering is infrastructure. Getting three different LLM providers to produce consistently structured outputs required way more iteration than building the UI.
· Stateless architecture is a superpower. No user data means no GDPR headaches, no privacy compliance cost, and nothing to maintain—exactly what a small team or solo dev wants.
What this is (and isn’t)
This isn’t a SaaS. It’s a fully functional demo platform that proves AI observability can be done without expensive tooling. If you’re evaluating LLM pipelines, debugging agent behavior, or just want a lightweight way to visualize system states, Cognilumin works right now.
🔗 Live demo: https://cognilumin.com
I’m genuinely looking for feedback—what would you add, what feels clunky, what would make you actually use this regularly? If someone sees a bigger vision for the domain + codebase, I’m open to that conversation too, but honestly I’d just love to hear what the dev community thinks.
Top comments (1)
The particle engine actually started as a mood‑visualization experiment, then I repurposed the math for system‑state mapping. Saved weeks
