Replay Showcase
https://www.replay.build/
Inspiration
I experienced the "Legacy Trap" firsthand at a 23,000-employee corporation. We had the budget to modernize, but we were paralyzed by the fear of breaking undocumented, 20-year-old backends. I realized that for legacy software, *the code is often useless or dangerous to touch. The only source of truth is the running application.*
We spend $300B annually on legacy modernization, yet 70% of projects fail because teams try to reverse-engineer systems by hand - watching old screens, writing specs, and hoping they caught everything.
I built Replay to automate this. The core insight is that UI is not static; it exists in time. Screenshots miss navigation, hover states, loading sequences, and error handling. *Video captures the full specification.*
What it does
Replay is a developer tool that turns screen recordings of legacy software (ERPs, CRMs, internal tools) into production-ready React code. It extracts:
*Visual Structure:* Layouts, components, and typography.
*Behavior:* Click paths, hover states, and user flows.
*Logic:* Input validation, error states, and loading sequences.
The output is a componentized codebase (React + Tailwind) and a living Design System, reducing migration time from months to days.
How we built it
Replay was built entirely by a solo founder using a cutting-edge AI-native workflow. The core engine relies on a complex multimodal pipeline powered by Google Gemini models:
1. *Gemini 3 Pro as "The Eyes":* We utilize the massive context window and multimodal capabilities of *Gemini 3 Pro* to process the raw video stream. It acts as the system's "eyes," perceiving not just static pixels, but the continuity of elements. It understands that a button press at 0:05 causes the modal at 0:07, capturing the temporal context that other models miss.
2. *Agentic Vision with Gemini 3 Flash:* To ensure enterprise-grade accuracy, we implemented a parallel **Agentic Vision layer using Gemini 3 Flash**. We call this "Sandwich Logic." While Pro generates the complex structure, Flash acts as a high-speed agent that verifies every step against the original video frames, hallucination-checking the output and ensuring data integrity.
3. *Logic Inference:* We prompt the models to infer business logic from visual cues (e.g., recognizing a red border as a "Validation Error" and generating the corresponding Zod schema).
Challenges we faced
*The "Static Trap":* Early iterations treated video as a series of screenshots, resulting in disjointed code. We had to engineer a pipeline that maintains state consistency across frames.
*Legacy Inconsistency:* Old software is visually messy (inconsistent fonts, mixed styles). We used Gemini 3 Pro to normalize these inputs into a clean, modern Tailwind-based Design System automatically.
Accomplishments that we're proud of
Building a fully functional video-to-code engine as a solo founder in under 2 months with 200 active useres and monetization.
Successfully running 2 enterprise pilots (Healthcare and Agency) where the tool is modernizing real-world legacy data.
Achieving a workflow where *Gemini 3 Pro and Flash work in tandem* to solve a problem that was previously considered "impossible to automate."
What's next for Replay
We are scaling the engine to handle multi-hour recording sessions and adding an more advanced "Agent Mode" that will allow developers to deploy the modernized code directly into their existing repositories. We aim to become the standard for safe, automated legacy migration.