# I built an AI observability platform with $0 – zero dependencies, zero ops, stateless

> Source: <https://dev.to/mamman_mohammed_86795ea0e/i-built-an-ai-observability-platform-with-0-zero-dependencies-zero-ops-stateless-4e7j>
> Published: 2026-07-29 20:31:35+00:00

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.

**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](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.