I built QasefAI - A bilingual AI Roast Card Generator powered by Gemini 2.5 Flash. Tweaked safety settings to BLOCK_NONE for raw sarcasm! 🔥
Hey everyone,
Wanted to share a side project I recently shipped that pushes Gemini's creative and sarcastic limits.
I built QasefAI (https://qasefai.lovable.app), a cyber/neon-themed web app that takes user traits or quiz answers and generates a brutal, witty "Roast Card" optimized for social media sharing.
Since Gemini is usually incredibly polite and helpful by default, here is how I tweaked the integration to give it some real attitude:
Technical Details & Gemini Setup:
- LLM Engine: Powered by `gemini-2.5-flash` for blazing-fast text generation and cost-efficient token usage.
- Bypassing the Politeness (Safety Settings): I dialed down the harassment and hate speech thresholds to `BLOCK_NONE` inside the server config. This completely unlocked Gemini's ability to generate raw, authentic internet-style sarcasm without getting censored or cut off mid-generation.
- Flawless Bilingual Sarcasm: Gemini automatically detects the input language. If English, it uses modern internet slang. If Arabic, it leverages highly localized, witty cultural humor (Levant/Gulf/Egyptian street humor). The linguistic adaptability is honestly impressive.
- Architecture: Built using Lovable (React/Vite). To ensure maximum security, the Gemini API call runs securely on the backend via TanStack Server Functions, keeping the API key completely safe from the client-side bundle.
Check it out here: https://qasefai.lovable.app
Give it a spin, drop your custom prompt or traits, and let me know how hard it roasted you! I’d love to get your feedback on how Gemini handles creative, edgy copywriting compared to other models. 🤙