ViralCoach

Inspiration

What it doesViralCoach: The AI Agent That "Sees" Your Content Strategy

💡 Inspiration We live in a creator economy where Distribution is King. Every minute, 500 hours of video are uploaded to YouTube. The barrier to entry for creation has dropped to zero thanks to Generative AI, but the barrier to attention has never been higher.

We realized that creators don't need another tool to write generic scripts. They need a Coach—an agent that can watch their video, see what their competitors are doing, and give mathematically grounded advice on how to stop being ignored.

🚀 What it does ViralCoach is a multimodal AI agent that acts as a 24/7 content strategist.

Multimodal Analysis: It watches your raw video files (visuals + audio), analyzing pacing, lighting, and energy. Competitor Roast: It live-scrapes your biggest competitors on TikTok using rtrvr.ai, extracting their hooks and view counts to benchmark your performance. Idea Forge: It takes random clips and stitches them into a cohesive viral narrative using proven retention frameworks. ⚙️ How we built it We built ViralCoach using a cutting-edge Agentic Stack:

Brain (Multimodal AI):

We used Google Gemini 1.5/2.0 Flash for its massive context window and native video understanding capabilities. It doesn't just read transcripts; it sees the video.

Eyes (Web Agents):

We integrated rtrvr.ai to perform real-time, deep web scraping of social media platforms (TikTok/Instagram) that are normally inaccessible to standard bots. Infrastructure: The app is built on Next.js 16 and Tailwind CSS v4 for a premium, responsive UI.

Workflow:

We utilized Prompt Driven Development (PDD), where an AI agent helped generate the codebase and test suites from high-level XML specifications. 🚧 Challenges we ran into The "Black Box" of Virality: Defining a "Viral Score" mathematically was hard. We had to iterate on our prompts to ensure the AI wasn't just hallucinating numbers but actually grounding them in visual signals (e.g., "The first 3 seconds are static"). Real-time Scraping: Social media platforms are notoriously hard to scrape. Using rtrvr.ai solved the access problem, but handling the unstructured data and mapping it to our schema required robust error handling and fallback logic.

Latency:

Multimodal analysis + Web Scraping = Slow. We optimized this by parallelizing requests and using optimistic UI states to keep the user engaged while the agents worked. 🏅 Accomplishments that we're proud of The "Roast" Feature: Successfully connecting a live web agent (rtrvr) to a reasoning engine (Gemini) to perform a real-time competitive analysis. It feels like magic when the AI tells you, "Competitor X got 1M views because they used a pattern interrupt, you used a static shot."

The UI:

We achieved a premium, "Dark Mode" aesthetic that feels like a professional SaaS tool, not a hackathon prototype.

Built With

• nextjs
• pdd
• rtrvr