← Back to field notes
2026-05-23

Why we built gemscout

Tier-1 funds chase the same 200 launches per quarter. Every Sequoia, Index, a16z scout watches the same Hacker News thread, the same YC batch, the same "exited stealth" PR loops. Meanwhile a founder in Krakow with 14 paying construction-PM customers, an AI radiology tool for Indian clinics with a 4.7 App Store rating, a Vietnamese fintech CRM with 30 SMB seats — they ship for 18 months without a single fund email landing in their inbox.

Those are the founders we want to back. So we built a pipeline.

The funnel, briefly

Every weekday morning a cron sources candidates across our active verticals — AI-native, open-source, bootstrapped OSS. We pull from where these founders actually live: Indie Hackers, MicroAcquire, App Store new launches in emerging markets, niche industry forums, regional X/LinkedIn signal. Not YC. Not "exited stealth today" PR.

Then the funnel does its job:

  • Layer 2 — light enrichment. Homepage reachable, founder names from /about, Wayback first-snapshot, GitHub 90-day commit velocity, App Store rating + count. Dead sites get cut here.
  • Layer 3 — stage filter. Is this actually pre-seed? Anything with a priced VC-led round or >$1M total funding gets dropped — they are no longer underground.
  • Layer 4 — deep research. Seven parallel sections of Perplexity research, Claude composite scoring under our underground-bootstrapper rubric. Composite < 3.0 auto-rejects.
  • Layer 5 — dossier. Claude Sonnet writes a markdown investor memo and we render it as a branded PDF (internal + external variants).

What survives lands in your feed. One dossier per vertical per weekday, on average.

What we are still figuring out

Three things keep us up.

Sourcing recall. The funnel works — composites we ship correlate with how the partner reads each founder by hand. But are we missing whole categories of underground bootstrapper because we are looking in the wrong places? Recent additions: Discord community size, Reddit subscriber counts, USPTO patent filings. More to come.

The quality bar. Layer 4's threshold is 3.0 for the underground rubric. Some weeks the funnel rejects everything — legitimately weak candidates — and the feed goes quiet. We are tuning sourcing volume (count: 10 per branch per run now, up from 5) and the surplus buffer so subscribers get a steady drip even on dry days. The bar itself stays.

The Founder's Pick. Once a month our partner hand-picks the strongest signal across all three live verticals and fully reveals it to every Free Preview subscriber. No paywall, no card. The dossier they would otherwise pay for, just for being on the list. The first one is up at /picks.

We will write more here as patterns emerge — sourcing experiments that work and ones that did not, retros on candidates we missed, the occasional market take. Bookmark this page or subscribe to a vertical's RSS feed.

— Artur (Arcada + TTL Capital)