Twitchy
Find Twitch creators who are growing, before the follower count says so.
The question I was chasing
Which creators are worth backing before everyone else has noticed them?
I am building Twitchy with a brand in the gaming space that wants to partner with creators on the way up, not established stars. The discovery tools they had all answered a different question. They rank on follower counts, views and engagement rates, which measure how big someone already is, not whether a partnership will mean anything. Paying for reach keeps the biggest number winning and the genuinely interesting creator missed. The catch is that the one thing that would answer it, who is actually growing, is the one thing Twitch does not keep. No past viewer counts, no follower growth, no hours streamed.
The constraints
Three rules pulled against each other. With no history and no location coming from the platform, momentum and geography had to be constructed rather than fetched. The system had to stay honest: no invented trend lines, no guessed countries, and a plain admission when there is not enough data to judge a window. And the enrichment layer, the off-platform research on every creator, had to cover a whole catalogue on a fixed budget.
The decisions that mattered
Own the history nobody keeps. Twitchy checks Twitch on a schedule and stores timestamped snapshots, so a picture of how a channel is actually moving builds up with calendar time.
Let the AI search with a fixed set of safe tools, never free access to the raw database. Only one of those tools can put creators on screen. What a brand sees is a reasoned shortlist, not a dump of rows.
Tie collection to search with one campaign brief. That brief decides who gets researched next and doubles as the brand's board layout, so what the system learns and what the client sees stay in lock-step. Adding a client is adding one brief, not rewiring the app.
Split research across three places. Live in the app for a search, on cloud-hosted helpers for web research that would time out in the app, and as a local job for bulk enrichment. That last one is a budgeted crawl, so enrichment fills the catalogue a band at a time.
Built with: Next.js, Supabase, Claude, a read-only MCP server, Vercel
Where it landed
A working system, in development with the client. Collection runs on a schedule and the agent search works.
Owning the history has a cost I underestimated. The job that turns stored snapshots into trends buckled the first time it met millions of rows. It had to be rebuilt to summarise them by the hour before it would finish at all. The system still does not read chat-level engagement, which is the thing that actually separates a community from an audience.
Part of the Rolling Waves work archive.