swarmFroge

You don't do the grunt work. You just Cmd / Approve. SwarmForge turns solo content creators (YouTube, Reels, TikTok, X) and independent media studios into high-velocity production powerhouses.

Comment

Inspiration

Solo creators, indie filmmakers, and digital production teams are essentially one-person movie studios today. Turning a raw 10-minute video concept or podcast into an engaging cinematic story, 3 viral vertical shorts, retention-engineered hooks, and a multi-platform distribution campaign requires 15+ hours of tedious manual grinding.

We asked ourselves: What if an autonomous studio lot of AI agents—a Showrunner, Scriptwriters, a ruthless Retention Critic, and a Compliance Supervisor—could co-produce blockbuster content alongside creators in real time?

That vision inspired SwarmForge: an autonomous multi-agent production studio that turns messy voice notes, video transcripts, and raw story concepts into cinema-grade production packs in seconds.

What it does

SwarmForge operates like an executive digital cinema backlot:

  1. Autonomous Showrunner Engine: Ingests raw voice dictations, script outlines, or transcripts, decomposing them into full-length YouTube scripts annotated with [B-ROLL CUES], scene framing directions, audio pacing notes, and a 3-act narrative structure.
  2. Multi-Format Repurposing Swarm: Parallel agent workers instantly generate 3 viral 9:16 vertical Shorts/Reels with 0–3 second hook engineering, visual text overlays, and a 7-post high-signal X thread.
  3. Adversarial Retention Critic: Evaluates generated screenplays against an adversarial scoring rubric (pacing, hook velocity, information density, and audience drop-off risk). Deliverables must pass an empirical threshold ($\ge 82/100$) before approval.
  4. Deterministic Safety & Compliance Gate: Zero-LLM grounding checks verify sponsor disclosures, brand safety, and factual claims without hallucinatory drift.
  5. Thumbnail & Hook Studio: Generates high-CTR cinematic thumbnail concepts, hook variations, and visual compositions with multi-aspect ratio previews (16:9, 9:16, 1:1).
  6. Executive Audio Briefing: Generates a 60-second synthesized Hollywood-style executive audio briefing summarizing the production strategy before export.

How we built it

SwarmForge was built using a modern, scalable full-stack agentic architecture:

  • AI & Reasoning Models: Google Gemini 3.7 Flash and Gemini 3.5 Flash for deep script reasoning, multi-character voice generation, and the adversarial critique loop; text-embedding-004 for semantic context and knowledge graph alignment.
  • Agent Orchestration: A multi-agent consensus pipeline coordinating a Lead Showrunner, Parallel Builders, and a Critic with autonomous state convergence.
  • Audio & Live Interaction: Web Audio API and Web Speech synthesis for real-time dictation, live frequency visualizers, and 60-second executive audio briefings.
  • Persistence & Cloud Infrastructure: Google Cloud Firestore for durable mission histories, prompt telemetry, and real-time state synchronization; Firebase Authentication for secure creator credentials.
  • Frontend & Visual Studio: React 19, TypeScript, TanStack Start/Router, Tailwind CSS v4, Motion layout animations, and HTML5 Canvas for real-time thumbnail compositing and teleprompter playback.

Challenges we ran into

  • Agent Consensus & Loop Convergence: Preventing multi-agent debates from spinning in infinite loops. We solved this by implementing strict iteration caps, dynamic temperature cooling, and deterministic milestone gates.
  • High-Fidelity Cinema Formatting: Ensuring LLM outputs respected structured cinematic script conventions ([B-ROLL CUE], scene action lines, dialogue margins) without breaking into raw unformatted text blocks. We created an intelligent client-side deliverable parser with multi-mode rendering (Formatted, Teleprompter, and Raw).
  • Sub-Second Audio Visualizer Synchronization: Synchronizing low-latency browser audio playback with dynamic waveform frequency bars during the executive audio briefing.

Accomplishments that we're proud of

  • True 4-Stage Autonomous Pipeline: Achieving an end-to-end flow from raw voice prompt to complete multi-deliverable media packs in under 15 seconds.
  • Deterministic Zero-LLM Guardrail: Enforcing strict regulatory and sponsor compliance through AST/regex validation before dispatching to external workspaces.
  • Integrated Teleprompter & Thumbnail Studio: Providing creators with a distraction-free recording teleprompter and real-time thumbnail hook tester directly inside the command center.
  • Full Cloud & Local Persistence: Seamless real-time sync across Firestore and local browser state with zero data loss.

What we learned

  • Adversarial critiques yield 3x better screenplays than single-prompt generation. Having a dedicated Critic agent push back on pacing and hook retention radically elevated the quality of the scripts.
  • Multi-agent collaboration works best when each agent has a strictly partitioned responsibility, typed schemas, and isolated telemetry logging.

What's next for SwarmForge

  • Direct timeline export integrations with Adobe Premiere Pro and DaVinci Resolve via OpenTimelineIO.
  • Real-time video frame rendering using text-to-video models (Veo / Imagen Video) to automatically generate B-roll clips directly from script cues.
  • Collaborative live multi-user writers' room where teams can co-pilot agents simultaneously over WebSockets.

Built With

  • cloud-data
  • gemini
  • gemini-3.5
  • goofle-auth
  • google-firebase
Share this project:

Updates

Submission history