About
How This Was Built
Methodology, architecture, and timeline from concept to deployment.
The Experiment
I treated the screener as a design challenge and a learning experience, not just a test to pass. The conventional answer might be to hire interns and build a manual pipeline. I asked a different question: what if the rote work didn't need to be done by humans at all? I built a proof of concept to find out.
The Method
Using my personal Claude Pro account, I sourced public domain educational videos with no existing metadata from the Internet Archive Commons — the same kind of raw video Khan Academy needs to process. I built a Claude-powered agentic pipeline that ingests a video, transcribes it, and generates all five required metadata outputs: SEO title, SEO description, timestamps, tags, and a thumbnail brief. I tested it against a single video. Total processing time: 10 minutes. Total cost: $0.02. I then built this website from scratch using Claude Code — a single prompt produced the full Next.js site, which I refined and deployed to Vercel. Total time from concept to live site: one day.
Why This Works
Four reasons this approach outperforms a manual intern pipeline:
Accurate — The video itself contains the SME's exact language, themes, and terminology. Claude generates metadata directly from that source material, not from guesswork.
Consistent — Every video goes through the same process. No variation in quality between interns, no training drift, no fatigue.
Cost-efficient — $0.02 per video. 1,000 videos costs approximately $20 in API fees. Compare that to three months of intern stipends.
Scalable — The same pipeline handles 1 video or 10,000. It can be extended to support localized regional content, multiple languages, and different content types with minimal modification.
The Architecture
This is not an argument against humans. Subject Matter Experts are irreplaceable — their knowledge, judgment, and voice are the product. What AI replaces is the rote, repetitive data entry that has nothing to do with subject matter expertise. The pipeline is designed so that every piece of AI-generated metadata passes through a human QA review before anything goes live. SMEs create. AI processes. Humans approve. Nothing publishes without sign-off.
Timeline