Twitchy
A creator-discovery platform for brands: find up-and-coming Twitch creators with genuine, growing communities, before they break out.
The question I was chasing
How do you find a creator worth backing before everyone else has already noticed them?
Most tools can't, because they are looking at the wrong thing. Discovery in the influencer economy runs on vanity metrics, follower counts, views, engagement rates, and those numbers are a weak guide to whether a creator genuinely matters or whether a partnership will mean anything. I wanted to find the up-and-coming ones, the people making something with real substance and a community that is genuinely growing around it, while they are still on the way up rather than after they have broken out.
Why it exists
This is being built with a real client: a brand in the gaming space that wanted to partner with creators on the way up rather than established stars. Underneath that sits the thing the whole platform is really arguing with. So much of the influencer economy runs on vanity metrics, paying for follower size on the assumption that reach equals value. It often does not, and it quietly props up a false economy where the biggest number wins and the genuinely interesting creator gets missed. Twitchy exists in part to step around that. The catch on the build side: Twitch keeps no history, no endpoint for a channel's past viewers, its follower growth or its hours streamed, so who is genuinely on the way up is the one question the platform will not answer.
The constraints
Three rules shaped the build, and they pulled against each other. Twitch exposes no history and no location, so momentum and geography had to be constructed rather than fetched. The system had to stay honest: never interpolate a trend or guess a country, and say so plainly when there is not yet enough data to judge a window. And the enrichment layer, the off-platform research on every creator, had to scale across a whole catalogue on a fixed budget rather than a runaway bill.
The decisions that mattered
Own the history nobody keeps. Since Twitch offers no past, Twitchy polls and stores timestamped snapshots on a schedule and builds its own time-series, so the data compounds with calendar time and you can actually watch a community grow rather than guess from a single number.
Put a read-only boundary between the model and the database. The search agent drives a small set of typed tools, never raw SQL, and only one of them, present a shortlist, can render cards. So what a brand sees is always the agent's reasoned verdict, not a raw dump of rows.
Tie collection to search with one config. A single campaign object decides who gets researched next and doubles as the brand's board preset, so what the system chooses to learn about and what it puts in front of the client stay in lock-step. Adding a new client is adding one object, not rewiring the app.
Run the research where it fits. The same Claude research runs three ways depending on the job: a live in-process loop for a search, cloud-hosted agents for the long web research that would time out a serverless function, and a local pipeline for enriching the catalogue in bulk on a fixed budget.
What it is
A creator-discovery platform for brands. Ask it in plain language, up-and-coming creators in a category with a small but genuinely engaged community, and it reasons over its own time-series like an analyst, looking for real momentum rather than a big follower count, and hands back a shortlist rather than a spreadsheet. The part I am most taken with is what sits on top: enrichment. For each creator it goes and learns their life beyond the stream, the other channels they post on, the other things they are into, where they actually are, any brands they have worked with. Suddenly they are not a gamer with a viewer count, they are a person with edges and interests, and a brand can find the ones whose other world genuinely fits, and make a far better call on a creative partnership. It all lands in one place.
Built with: Next.js / Supabase / Claude / a read-only MCP server / Vercel
Where it landed
A working system, in development with the client. Ingestion runs on a schedule, the agent search works, and enrichment is filling the catalogue band by band. The honest edges are the interesting part. The metrics rollup buckled the first time it met millions of snapshot rows, and had to be rebuilt to downsample into hourly buckets before it would finish at all. Enrichment still covers a fraction of the catalogue, a deliberate budgeted crawl rather than a switch you flip. And the truest signal, real chat-level engagement, the thing that actually separates a community from an audience, is the next thing to add. It is the clearest thing I have built about where my work is heading: the commercial read and the engineering in one pair of hands.
Part of the Rolling Waves work archive.