SparkStudio AI

💡 Inspiration

Content creators spend 80% of their time on repetitive, tedious pre-production tasks—scriptwriting, keyword research, storyboarding, rendering, and captioning—leaving only 20% for actual creative work. Existing AI tools are highly fragmented: you use ChatGPT for scripts, Midjourney for visuals, CapCut for captions, and manually stitch them together with no cohesive, connected workflow.

We built SparkStudio AI to solve this. It is a production-grade, stateful multi-agent harness that connects specialized AI agents into a single, cohesive, self-healing pipeline. One topic, one click, complete video production kit in under 2 minutes.

⚙️ What it does

SparkStudio AI orchestrates 9 specialized AI agents to work collaboratively on a content pipeline:

🔍 Research Agent: Performs audience analysis, extracts viral angles, and generates high-impact hooks.
✍️ Script Agent: Writes the complete spoken script with precise timestamps and visual cues.
🎬 Storyboard Agent: Generates highly descriptive scene-by-scene prompts.
🎨 Visual Generator: Integrates with Hugging Face FLUX to render high-adherence storyboard stills.
🖼️ Thumbnail Agent: Recommends click-worthy thumbnail concepts.
📈 SEO Agent: Researches top keywords and description copy for YouTube/TikTok algorithms.
📝 Subtitle Agent: Renders synchronized .srt and .vtt caption files.
🎙️ Voice Agent: Generates descriptive speech instructions.
🏆 Quality Auditor: Evaluates readability, pacing, and retention scores, giving optimization suggestions.

The harness streams the agents' thoughts to the user in real-time, compiles a ZIP package containing your script, subtitles, rendered visual stills, and audit reports, and provides a one-click download.

🏗️ How we built it

We engineered a robust full-stack architecture built to scale:

Frontend: Next.js 14, TypeScript, TailwindCSS, and Framer Motion for modern, responsive dashboard layouts and fluid micro-animations.
Backend: FastAPI (Python) for asynchronous agent orchestration and Server-Sent Events (SSE) for streaming real-time log logs.
Database & State: Supabase (PostgreSQL) for session management and project persistence.
Inference Engines: Groq API (LLaMA 3) for blazing-fast text inference (~800 tokens/sec) and the Hugging Face SDK (FLUX.1-schnell) for image generation.

🧠 Challenges we faced

Context Decay: Passing entire chat histories between 9 agents causes context drift and hits API token limits. We solved this by building a custom Context Compactor that serializes only structured JSON summaries to downstream agents.
API Rate Limits: Free-tier inference keys return frequent 429/503 rate limits when hit concurrently. We implemented a robust Self-Healing layer featuring asynchronous retry loops with exponential backoff and silent fallback providers (like Pollinations).
Database Caching: Initial image generation would regenerate every time a user visited the tab. We linked backend generation directly with Supabase to cache raw base64 data URLs, enabling instant reloads.

🏆 Accomplishments that we're proud of

300x Workflow Speedup: Replaced a manual pre-production workflow that typically takes 10+ hours with a 2-minute autonomous pipeline.
Flawless Real-Time Streaming: The Server-Sent Events channel works perfectly, allowing users to watch agents think, debug, and cooperate live.
Durable Persistence: A completed project can be re-accessed at any time, with all storyboard visuals, scripts, and audits loading instantly from Supabase.

📚 What we learned

We gained hands-on experience in advanced multi-agent design patterns, state management in long-horizon pipelines, and optimization techniques for serverless image generation over high-latency networks.

🚀 What's next for SparkStudio AI

Auto-Stitching Engine: Integrating FFmpeg on the backend to automatically combine the generated script, voiceover, and storyboard images into a draft video format.
Publishing Integrations: Direct one-click publishing pipelines to YouTube Shorts, TikTok, and Instagram Reels APIs.