← sources

Day 4: Building an AI Website Auditor | $1,500 Budget | $7 Spent | Working on the AI's Evidence Pipeline

source: reddit (reddit:AI_Agents)published 2026-11-03 04:26ingested 2026-11-04 02:36

retained text

After a few days away for travel, work, and some serious jet lag, I'm back to testing. (BTW - I'm a UX/UI designer building an AI-powered website auditor with a CA$1,500 budget, no technical cofounder, and ChatGPT helping me figure out the backend.)

The goal is to create an AI auditor that evaluates web pages using a structured UX framework, rather than generating a list of generic recommendations.

I'm starting with UX only, one page at a time.

**What I've tested so far:**

* **Firecrawl:** Extracts webpage content and HTML, but doesn't necessarily capture the complete interactive experience.
* **Playwright:** Successfully set up locally to interact with JavaScript-driven navigation and collect screenshots and page information.
* **Evidence collection:** Generated structured page data that can eventually be passed to an AI model alongside screenshots and business context.

The system will use three inputs: a URL, a user-provided full-page screenshot, and a short business description covering the audience and the page's purpose.

**The challenge I'm working through:**

How do you give an AI enough reliable context to evaluate a website without overwhelming it with irrelevant data?

HTML alone isn't enough. Screenshots alone aren't enough. And extracting every possible element could become expensive.

I already have a UX evaluation framework developed from my design experience. The next step is to combine that framework with the collected evidence and test whether the model can produce specific, evidence-backed findings.

**Budget update:**

* Total budget: CA$1,500
* Spent: $7 on research
* Backend and infrastructure: $0

I'm still in the experimentation stage. No working AI audit engine yet.

**Question for AI builders:** Would you preprocess and filter the extracted website evidence before passing it to an LLM, or let the model work with a larger raw dataset? I'm particularly interested in balancing context quality, token costs, and reliability.

document fetched 2026-11-04 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xmjf5r/day_4_building_an_ai_website_auditor_1500_budget