Ask questions about the startup curriculum
The short version: I hate watching videos.
Elevate Ventures has an excellent startup curriculum — 39 videos covering everything from fundraising to hardtech manufacturing to go-to-market strategy. The content is genuinely useful for Indiana founders.
But it's locked inside hours of video. If you want to know what they say about term sheets, you have to scrub through recordings hoping to find the right 90 seconds. Nobody does that. The content sits unwatched.
I'm a reader, not a watcher. I process information faster as text. And if I want to find something specific, I search — I don't scroll through a video timeline.
This makes video content a perfect candidate for RAG (retrieval-augmented generation). Transcribe the videos, chunk the transcripts, embed them into vectors, and suddenly you can ask questions about the content and get answers with precise citations back to the exact moment in the source video.
The video doesn't go away — it becomes a citation you click when you want the full context. Text-first, video-on-demand.
I wanted to show that a useful AI-powered app can be built entirely on Cloudflare's edge stack — no servers, no Docker, no managed databases, no OpenAI API keys. Workers AI + Vectorize + D1 + R2, all running at the edge, all on one platform. The whole thing deploys with a single wrangler deploy.
This is a proof of concept for CTK Advisors: any organization with video or document libraries can make that content searchable and conversational with the same architecture.
A RAG-powered chat interface over Elevate Ventures' startup learning video series — entirely on Cloudflare's edge with zero servers to manage.
| Layer | Choice | Why |
|---|---|---|
| Transcription | MLX Whisper (large-v3-turbo) | Runs natively on Apple Silicon, word-level timestamps for precise chunking |
| Chunking | ~500 tokens, 50-token overlap | Balances context richness with embedding precision |
| Embeddings | bge-base-en-v1.5 (768-dim) | Best-in-class open embedding model on CF Workers AI |
| LLM | llama-3.3-70b-instruct-fp8-fast | Strong instruction-following, runs at edge with no cold start |
| Database | Cloudflare D1 (SQLite at edge) | Relational metadata co-located with worker, zero latency |
| Vector Store | Cloudflare Vectorize | Native integration, no external vector DB needed |
| Hosting | Cloudflare Workers + Assets | Global edge deployment, static + API in one worker |
| Video | Vimeo embed with #t= seek | Leverages existing Elevate Ventures Vimeo hosting |
Everything runs on Cloudflare's edge — no servers, no Docker, no databases to manage
Embedding + vector search + LLM all happen at the edge
Every answer links back to exact video timestamps — click a citation and the player seeks to that moment
Full Elevate Learning curriculum, chunked with word-level timestamp alignment from Whisper