Recently, I built an automated data-analysis and comparison directory for board games called Tabletop Dev Labs .
The core idea is simple: instead of forcing casual players to dig through thousands of dense forum posts, a Python pipeline scrapes metadata and community reviews, feeds them into an LLM to extract structured JSON metrics, and builds a fast static site (SSG) using Astro deployed on Cloudflare Pages .
One of the flagship features of the site is a Luck vs. Skill Balance metric (scored from 0 to 100) alongside a 5-axis radar chart ( Brain Burn, Setup Hassle, Friendship Destroyer, Party Vibe, Replayability ).
However, after generating my first ~80 pages using a local LLM, I ran into a frustrating problem that almost broke my comparison pages: almost every single game was being rated 50% Luck / 50% Skill.
Here is what happened, why it broke my comparison pages, and how migrating the evaluation pipeline to Gemini 3.5 Flash-Lite completely solved it.
The Architecture
To keep hosting costs at $0,…