RealStateVideo is a live SaaS that turns real-estate property listings into platform-optimized marketing videos. It handles the full pipeline: vision-based property analysis, script generation with viral-hook scoring, text-to-speech narration, word-level caption alignment, automated FFmpeg rendering, and a quality-evaluation step that gates bad renders before delivery.
The Problem
Real-estate agents need short marketing videos for every listing: for Instagram, TikTok, YouTube Shorts, and property portals. Producing these manually takes 30 to 60 minutes per video (script, record voiceover, edit, add captions, export). Agents have dozens of listings. The math doesn’t work.
The Approach
Vision analysis takes property photos and extracts key features (rooms, square footage, highlights) using a multimodal LLM.
Script generation writes a video script optimized for the target platform (vertical for TikTok, landscape for YouTube). A viral-hook scoring step evaluates the opening line and regenerates if the score is below threshold.
TTS + caption alignment generates narration via text-to-speech and aligns captions at the word level for precise subtitle timing.
FFmpeg rendering composites the video: property photos with Ken Burns effect, captions, platform-specific aspect ratio, and branding. Fully automated, no manual editing.
Quality eval runs an LLM assessment on the final render: checks audio sync, caption readability, script coherence, and overall quality. Renders that don’t pass are flagged for regeneration.
Stripe billing handles subscriptions. Public REST API for programmatic access.
Key Decisions
Quality gate before delivery. Most AI video tools ship whatever the model produces. A quality-eval step catches bad renders (audio out of sync, gibberish captions, incoherent scripts) before they reach the user. This trades a small amount of latency for significantly higher output quality.
Platform-specific generation. A TikTok video has different requirements than a YouTube video: different aspect ratio, different script style, different hook strategy. Generating platform-specific output (rather than one-size-fits-all) dramatically improves engagement.
Trigger.dev for pipeline orchestration. The video generation pipeline has 6+ sequential steps with different failure modes. Trigger.dev handles retries, timeouts, and job monitoring.
Results
- Automated end-to-end. Listing photos in, marketing video out
- Platform-optimized output for Instagram, TikTok, YouTube
- Quality-gated renders (bad videos don’t ship)
- Live SaaS with Stripe billing and public API
- Reduces per-video production time from 30 to 60 minutes to under 5
Retrospective
The quality gate was the highest-ROI decision. It turned inconsistent AI output into reliably good videos. Platform-specific generation was also key (agents noticed the difference immediately). If rebuilding, I’d add more template variety and a direct integration with listing platforms (MLS import) to eliminate the manual photo-upload step.