🌲 Inspiration

My best friend Ranju spent five years building "PureVeda,"(name changed) a digital sanctuary for holistic wellness, Ayurveda, and toxin-free living. Her readers came for serenity, not sales pitches. But as PureVeda grew, she faced the "AdSense Paradox." To keep the lights on, she turned on Google ads, trusting their AI to curate the space.

The Erosion (The Vibe Clash): The nightmare began last year in an August summer day. A reader messaged her: "Ranju, I thought you were against chemical shortcuts?" Ranju opened her site and felt a physical jolt of nausea. Nestled between her article on The Healing Power of Turmeric was a neon-flashing ad for "Weight Loss Injections" featuring a grotesque, hyper-saturated "before and after" photo. A few scrolls later, a "Toenail Fungus" ad appeared next to her recipe for organic moon milk.

Google’s AI saw these as "Safe" (Health category), but they were Brand Toxic. The calm, earthy vibe of PureVeda was being replaced by the "Clickbait Aesthetic." Her high-end audience started to bounce, and her newsletter unsubscribes spiked by 14% in one month. Her monthly revenue plummeted.

The Manual Prison (The Review Hell): Determined to save her brand, Ranju dove into the AdSense Ad Review Center. She found herself in a digital labor camp.

The Scale: 842 ads were waiting for review.

The Fatigue: She had to click each one, wait for the preview to load, and hit "Block."

The Futility: She’d spend three hours blocking "Miracle Cures," only for the same advertisers to use different URLs five minutes later. Ranju was no longer a wellness expert; she was an unpaid, exhausted janitor.


⚙️ What it does

PubAdVibe is the first Aesthetic Intelligence Layer for the premium web, powered by the multimodal reasoning of Gemini 3 Pro. While traditional filters only catch malware, PubAdVibe audits the 'soul' of an advertisement. Thereby promoting - "Ethical AI for better web experiences."

PubAdVibe is the "Private Security" Ranju never had. Instead of Ranju manually clicking, PubAdVibe uses Gemini 3 Pro’s "High Thinking" Mode to act as a Brand Concierge.

Semantic Vibe-Check: PubAdVibe doesn't just look at "categories." It reads Ranju's "Brand Manifesto" (Minimalist, Earthy, Organic). It sees a neon weight-loss ad and rejects it instantly because the Visual Sentiment clashes with the site's CSS and tone.

Deceptive UI Shield: It catches the "Fake X" buttons and "System Warning" ads that try to trick her older readers—the ones that Google's bots miss because the code is technically "clean."

Agentic Peace: Ranju sets her "Vibe Score" to 85. PubAdVibe automatically blocks anything lower. She gets her Sundays back. PureVeda smells like sandalwood again, not clickbait.

PubAdVibe acts as a visual firewall. It ingests ad creative assets and uses Gemini 3 Pro to perform a three-tier audit:

  1. Aesthetic Harmony: Does this ad fit the brand's "Vibe Manifesto" (e.g., minimalist vs. aggressive)?
  2. Deceptive UI Detection: It "looks" for fake 'X' buttons or system dialogs embedded in images that standard scanners miss.
  3. Policy Guard (RAF 2026): It scans for synthetic deepfake endorsements and fraudulent "get rich quick" visual tropes.

Architecture: https://github.com/San1973/PubAdVibe/blob/main/architecture.md

"PubAdVibe doesn't just block ads; it curates trust."


🧱 How we built it

We built a modern, premium stack using React 19 and Tailwind CSS, centered around the Gemini 3 Pro Preview model. We leveraged Gemini’s strict JSON output capabilities to turn subjective "vibes" into objective data points. By using a custom-engineered prompt that injects a site's "Brand Manifesto" as context, we turned a generic LLM into a specialized Brand Safety Officer.


🧗 Challenges we ran into

Defining "vibes" in a way an AI can quantify was our greatest hurdle. We had to iterate on prompt engineering to ensure the model didn't just flag ads for being "ugly," but specifically for "violating brand-specific constraints." We also tackled the technical challenge of real-time multimodal ingestion, ensuring high-resolution image analysis didn't compromise the "low-latency" feel of the security dashboard.


🏅 Accomplishments that we're proud of

  • We are incredibly proud of our Deceptive UI Heuristic. During testing, PubAdVibe successfully identified "pixel-traps" (fake close buttons) that passed traditional code-based security audits. Seeing the "Revenue Protected" counter tick up as the AI "purged" visual toxins felt like a true milestone for a cleaner web.

  • Novelty benchmark: excelled far better when we are comparing PubAdVibe to existing ad blockers, which highlights Gemini 3's semantic depth (e.g., detecting "soul-violating" subtleties humans miss). Link: https://www.youtube.com/playlist?list=PL47oafi_Pj7NZD_EDqlvGmlYa53jGeJjC


📚 What we learned

We learned that the future of web security isn't just in the code—it's in the visual experience. Code analysis is no longer enough to protect users from sophisticated AI-generated scams. To protect the web, we need AI that can "see" and "reason" like a human, but with the speed of a machine.


🔭 What's next for PubAdVibe

  • Our next step is the Gateway Proxy: a server-side implementation that intercepts AdSense and Header Bidding responses at the edge, performing the Vibe-Check before the ad ever hits the user's browser. We envision a web where every premium publisher has an "Aesthetic Intelligence" layer protecting their soul and their revenue.

  • Enhance architecture: Adding proxy integration for real-time ad interception (e.g., via browser extension, which is already WIP). using JSON outputs for scoring (e.g., "aggressiveness: 7/10"). we plan to host on Vercel with fallback for API limits.

  • Extend to "adaptive vibes" (e.g., user feedback loops where AI refines manifesto via multi-turn reasoning). We also plan to add fun elements like visualized "vibe heatmaps" (Gemini-generated overlays on ads), in the near future.

  • We plan to add one "amaze" feature, if AI tech is sponsored by Google: AI-generated "alternative ad" suggestions (e.g., remix intrusive ads into serene versions using text-to-image), thereby "personalizing" ads as per The-Vibe

  • Integrating multi-agent setup, using Gemini 3's real strength: One agent for visual analysis, another for semantic vibe-check, a third for user personalization (e.g., adapt based on site history). Then we can "Watch AI debate ad ethics in real-time".


🚀 PubAdVibe: Protecting the soul of the premium web.

Built With

Share this project:

Updates