How we built a 100% automated, bilingual AI video pipeline for <$0.50 (Gemini 3.8 + MDD + RunPod + Blender MCP)

Hey Lounge builders,

Like many solo founders building in public, I hit a massive bottleneck last week: creating high-converting, vertical video content (Reels/Shorts/TikTok/X) consistently without spending 8 hours inside CapCut or paying $1,000s/month to agencies.

Most off-the-shelf "AI video generators" either look like cheap slideshows (Ken Burns zoom), hallucinate scrambled gibberish whenever there's a logo or UI text, or lock you into expensive monthly SaaS seats.

So we engineered our own autonomous, end-to-end video production pipeline. Here is the exact architecture we shipped:

---

Everything begins inside Google Antigravity paired with Gemini 3.8 Flash acting as the Systems Architect. Instead of starting from zero context, the agent taps into our Memory-Driven Development (MDD) framework—a versioned external memory architecture we open-sourced (github.com/otavioserra/conn2flow-ai-workspace). It already understands our product (Conn2Flow 3.0), audience pain points, tone, and past changelogs.

• Generates high-retention 60s bilingual scripts (English & Brazilian Portuguese).
• Prompts 9 native vertical (9:16) photorealistic keyframe scenes.

• Fast Python neural TTS synthesizes two clean studio-grade voice tracks (American English + Brazilian Portuguese) in under 3 seconds.
• Local Faster-Whisper on our workstation extracts word-level timestamps (word_timestamps=True), giving us the exact millisecond when each word and sentence begins and ends.

Instead of cooking our local GPU or paying flat SaaS fees:

• A custom CLI (runpod_gpu.py) provisions an on-demand NVIDIA RTX A5000 (24GB VRAM) for $0.27/hr.
• Includes a standalone local watchdog script (guard) with a hard deadline timer—if anything hangs, it automatically terminates the pod so runaway cloud bills are mathematically impossible.
• Inside the pod, ComfyUI runs Wan 14B GGUF (High/Low Noise) to animate the 9 vertical keyframes with true fluid dynamics, lighting, and camera physics (not static pans).

Diffusion models are notorious for butchering text, buttons, and logos across frames. Our solution: The keyframes are generated 100% textless.

• Our agent hooks directly into Blender 5.2 VSE via an MCP bridge (Model Context Protocol).
• Blender imports the raw Wan 14B video clips, aligns them to the Whisper speech timeline, and composites crisp, native 3D typography, vector logos, and animated dynamic subtitles directly over the video tracks!
• Renders the final 1080x1920 masters in both English and Portuguese, accompanied by an original synth score.

---

• Cost per complete 60s video (9 high-end cinematic clips): ~$0.35 to $0.50 (RunPod compute time).
• Setup Overhead: Diluted by batching 3 to 4 video campaigns per pod spin-up.
• Human Touch Needed: Review the concept, approve the renders, and publish.

The entire system runs on reproducible scripts, open-source models, and MCP tool automation.

I’m attaching our latest 60s vertical render below. Would love your feedback on the pacing, and happy to dive deep into any part of the stack (ComfyUI GGUF setup, RunPod watchdog, or Blender MCP scripting) with fellow builders!

What does your current marketing video workflow look like?

Video post: https://lnkd.in/p/eSd7j3_b