One video in. A full week of content out.
Every creator has the same problem: you spend hours making one great video, and turning it into a TikTok, a tweet thread, and a LinkedIn post is another hour of work-- PER PLATFORM. Cascade is a web app that takes a single source video and, through real video processing, word-level speech-to-text, and multi-stage AI prompting, automatically produces a full week of platform-native content: a captioned short-form clip with AI-sourced b-roll, plus a Twitter/X thread, a LinkedIn post, an Instagram caption, and a TikTok/Reels script.
Every creator has the same problem: you spend hours making one great video, and turning it into a TikTok, a tweet thread, and a LinkedIn post is another hour of work-- PER PLATFORM. Cascade is a web app that takes a single source video and, through real video processing, word-level speech-to-text, and multi-stage AI prompting, automatically produces a full week of platform-native content: a captioned short-form clip with AI-sourced b-roll, plus a Twitter/X thread, a LinkedIn post, an Instagram caption, and a TikTok/Reels script.
- Video Ingestion Pipeline: Accepts an uploaded source video and extracts audio for processing.
- Word-Level Transcription: Speech-to-text with word-level timestamps, powering both captions and clip selection.
- Multi-Stage AI Prompting: Structured, validated LLM outputs that (1) select the best short-form moment, (2) write platform-native posts, and (3) generate a shot/caption plan.
- Automated Clip Assembly: FFmpeg-driven cutting, karaoke-style burned-in captions (ASS subtitle format), and AI-sourced b-roll compositing (Pexels API).
- Platform Post Generation: Auto-generated Twitter/X thread, LinkedIn post, Instagram caption, and TikTok/Reels script from the same source video.
- Job Status Dashboard: Async job pipeline (upload -> processing -> done) with live progress polling on the frontend.
- Multi-Clip Selection: Detect and offer several strong candidate moments instead of just one.
- Editable Output: Let users tweak generated captions/posts before export instead of only copy-paste.
- Style/Tone Presets: Let users pick a voice (professional, funny, dramatic) that shifts all generated text together.
- Direct Publishing: OAuth into platforms and post directly instead of copy/paste.
- YouTube Link Ingestion: Accept a YouTube URL (via yt-dlp) as an alternative to a manual upload.
Week 1: The Setup 📁
- General:
- Decide team roles (Frontend vs Backend)
- Set up GitHub repo,
.envtemplate, and API keys (transcription API, LLM, Pexels) - Everyone installs the stack and posts a working "Hello World" screenshot in the team chat (lol)
- Read FFmpeg and yt-dlp docs now and familiarize yourself
- Frontend:
- Finalize UI design and UX layout (Figma)
- Backend:
- Make test requests to the transcription API and the LLM (confirm structured/JSON output works)
Week 2: Hello World ⚡
- Frontend:
- Initialize React (Vite) + TypeScript + Tailwind project
- Build basic layout: upload area, job status area, results area (stubbed)
- Backend:
- Stand up FastAPI skeleton with an in-memory job dictionary
- Build a fake background task that sleeps, updates progress a couple of times, then marks a job "done" with a fake result
- Confirm the frontend can POST a job and poll
GET /jobs/{id}through the fake stages end to end
Week 3: Real Ingestion 🎬
- Backend:
- Replace the fake job with real video upload handling
- Extract audio via FFmpeg/FFprobe (subprocess, explicit CLI flags)
- Call the transcription API and store word-level timestamps
- Frontend:
- Wire the upload form to the real endpoint
- Show real (not fake) progress states as they come back
Week 4: The Brain 🧠
- Backend:
- Prompt engineering: summarize the transcript into a brief
- Multi-stage prompting with structured outputs: best-moment selection, social post generation, caption/shot plan
- Add output validation (retry or flag malformed structured responses)
- Frontend:
- Build UI components to display generated posts once available
Week 5: The Cut 🎞️
- Backend:
- Cut the selected clip with FFmpeg based on the chosen timestamps
- Burn in word-synced captions using ASS karaoke tags
- Pull and composite b-roll from the Pexels Videos API
- Frontend:
- Style the "generated clip" preview area (native HTML5 video player)
Week 6: Full Loop 🔄
- Backend:
- Connect every stage (ingest -> transcribe -> prompt -> cut -> caption -> b-roll) into one real end-to-end pipeline
- Return the finished video + all four platform posts from a single job
- Frontend:
- Full integration: real upload -> real polling -> real results, no more stubs/fakes anywhere
Week 7: Everything Works ✅
- All:
- Run the full pipeline against every test video the team has, back to back-- fix whatever breaks
- NO NEW FUNCTIONALITY AFTER TONIGHT!!!!
- Add timeouts/retries on every external API call, clear error messages (not generic "something went wrong"), transcript caching by file hash, and temp file cleanup after each job
Week 8: Polish & Reliability + Presentation Creation ✨
- Frontend: Polish fonts/colors/spacing, fix layout bugs, add a "How it Works" section
- Backend: Review API usage/quota limits, fix edge cases found in testing, final code review
- All: Confirm the demo path (specific test video -> known-good output) is bulletproof, WORK ON PRESENTATION!!!
Week 9: Presentation Day Prep 🎬
- Record a backup demo video in case live generation fails on stage
- Practice the pitch and live walkthrough, timed
- Freeze the repo--no last-minute changes the night before presentation
IDE: VSCode
Version Control: Git / GitHub
Design: Figma
Frontend
- Framework: React (Vite) + TypeScript
- Styling: Tailwind CSS
- State Management: Zustand
- Icons: Lucide React
- Playback: Native HTML5
<video>element
Backend & AI
- Runtime/Framework: Python + FastAPI
- Async Jobs: FastAPI BackgroundTasks + in-memory job dictionary
- Validation: Pydantic
- Media Processing: FFmpeg / FFprobe (subprocess, explicit CLI flags)
- Captions: ASS subtitle format with karaoke tags for word-synced highlighting
- Transcription: Hosted speech-to-text API with word-level timestamp granularity
- LLM: OpenAI or Google Gemini (Flash)-- structured outputs, multi-stage prompting
- Stock Footage: Pexels Videos API (free tier: 200 req/hour, 20,000/month)
- Storage: Local filesystem (S3 if time allows)
Deployment
- Local for the demo
- If hosted: a container-based host that supports long-running processes and FFmpeg (Render, Railway, Fly.io)-- not Vercel/Netlify, which time out before a render finishes and don't ship FFmpeg
Frontend:
- React Documentation
- Vite Guide
- Tailwind CSS Docs
- Zustand State Management
- 🎥 React + Vite + TypeScript + Tailwind Crash Course (YouTube)
- 🎥 Tailwind CSS Full Course 2026 (YouTube)
- 🎥 Tailwind CSS Crash Course, Traversy Media (YouTube)
- 🎥 Zustand Beginner Tutorial (YouTube)
- 🎥 5 Zustand Best Practices in 5 Minutes (YouTube)
Backend:
- FastAPI Tutorial
- FastAPI Background Tasks
- Pydantic Docs
- 🎥 FastAPI Full Crash Course (YouTube, freeCodeCamp)
- 🎥 FastAPI Full Crash Course, Python's Fastest Web Framework (YouTube)
- 🎥 Python Pydantic Tutorial: Complete Data Validation Course, Corey Schafer (YouTube)
Media Processing:
- FFmpeg Documentation
- FFmpeg Filters
- FFmpeg Wiki
- ASS Subtitle Tags (karaoke/word highlighting)
- yt-dlp
- 🎥 FFmpeg Tutorial for Beginners: Mastering the Command Line (YouTube)
- 📝 Burning ASS Subtitles with FFmpeg — subtitles filter reference (written guide)
- 📝 How to Add Subtitles to a Video with FFmpeg, incl. ASS format (written guide, Bannerbear)
AI & APIs:
- OpenAI Speech-to-Text
- OpenAI Structured Outputs
- Google Gemini API Docs
- Pexels API Documentation
- 🎥 OpenAI Structured Output Tutorial (YouTube)
- 🎥 Gemini API with Python — Getting Started (YouTube)
Software to Install:
- Abhishri Dave
- Fawaz Jamshed
- Kashmine Zadia Shaik
- Manila Cholleti
- Sai Varun Mallela
Project Manager: Mohammad Mehrab
Industry Mentor: Jackson Parrish