Inspiration
Every creator we talked to said the same thing: filming and editing is only half the job. The other half — writing a title, tags, and description that actually get found — eats time and is mostly guesswork. Most people either copy what a competitor did or throw the topic at ChatGPT and hope for the best, with no real connection to what's actually ranking. We wanted to fix that guesswork problem directly.
What it does
TagGenie is an AI SEO copilot for YouTube creators. Give it a video topic (or the URL of a video you've already uploaded), and it:
Researches — searches YouTube for the top-performing videos on that topic and pulls their tags and title patterns Grounds — aggregates that into a keyword signal: what's actually working right now in that niche, not what worked a year ago Generates — feeds that real signal (plus an optional transcript) to an LLM, which returns 5 title options, 15–20 tags, and a hook-first description, all ready to paste
How we built it
The core is a Streamlit app for a fast, single-file UI with no frontend build step — important on a hackathon timebox. Under the hood:
YouTube Data API v3 for competitor search and video metadata (read-only, no OAuth needed) youtube-transcript-api to pull transcripts for existing videos Groq (Llama 3.3 70B) for generation, using a strict-JSON prompt so the model's output maps directly into the UI's copy-paste boxes
We deliberately kept the AI step downstream of the research step — the model never generates from a blank slate, it always has real competitor data to work from.
Challenges we ran into
Getting from "runs on my machine" to "runs anywhere" surfaced a few real bugs: dependency version mismatches between the LLM SDK and its underlying HTTP library caused cryptic unexpected keyword argument errors, and we hit a couple of API auth walls (invalid keys, low credit balances) before landing on a fully free stack. We also had to swap our original LLM provider mid-build once we realized a truly free, no-credit-card option mattered more for accessibility than sticking with our first pick.
Accomplishments that we're proud of
A working end-to-end pipeline — real API calls, real data, real generation, not a mockup Fully free to run: no credit card, no paid tier, on either API Metadata that's actually grounded in live competitor data instead of AI guesswork Deployable as a public website in minutes via Streamlit Community Cloud
What we learned
Grounding matters more than model size. A smaller, faster model fed real competitor keyword data consistently outperformed a "just ask the AI" approach — the research step is doing more work than the generation step. We also learned firsthand how fragile AI SDK dependency chains can be, and how much smoother a build goes once you settle on a free-tier stack from the start instead of retrofitting it later.
What's next for TagGenie
One-click Apply — push the chosen title/tags/description straight to YouTube via OAuth and videos.update A/B title testing — use the YouTube Analytics API to see which generated title actually won after publishing Thumbnail text suggestions alongside each title option Batch mode — run the whole pipeline across a channel's entire backlog in one pass
Log in or sign up for Devpost to join the conversation.