frontrun's whole product reduces to one loop: watch who venture investors follow on X, diff it continuously, and timestamp the moment several converge on the same company. simple to say; the edge is in running it well for a long time. this is the data story, with numbers.

the numbers

• 1,000+ tracked discovery accounts - investors, founders, and operators whose follows are load-bearing. curation matters more than volume: one gp with a fresh fund outweighs a hundred anon accounts.
• 42,000+ companies in the catalog, each with sector classification and description embeddings.
• every company carries provenance: first_flagged_at (when tracked investors first converged) and followers-at-flag. that's what makes "we saw it early" checkable instead of vibes.
• the signal is a diff, not a snapshot. a follow list today is trivia; the change between yesterday and today, across a thousand curated lists, is deal flow.

why follows leak information

investors research before they invest, and while researching, they follow. it's near-zero-cost behavior with no announcement attached - which is exactly why it's honest. our favorite receipts:

• @techdollarhq - convergence at 13 followers, feb 23. the $3M pre-seed went public 123 days later.
• @orthogonal_sh - flagged 184 days before a $4.3M round led by Pantera.
• @rialto_xyz - flagged jun 5 at 18 followers; 26 days later it was announced as a Robinhood chain launch partner. fundraises aren't the only thing that leaks - partnerships do too.
• @pascaldottrade - flagged dec 27, back when they were still pascal_markets. before the rebrand, before the invite codes.

the unglamorous 80%: junk filtering

raw follow data is mostly noise. the pipeline that makes it usable:

• entity classification - is this account a company, a person, a community? (llm-classified, human-corrected, corrections stick.)
• junk suppression - nft mints, meme tokens, engagement-farm accounts. an early lesson: without an aggressive junk filter, one airdrop season poisons a month of reports.
• cross-report dedup - a company appears in your report once when it's net-new, not every day it exists.
• follower-quality weighting - whose follow is it? a fund gp's follow and a bot's follow are not the same event.

what we got wrong along the way

honest list: we shipped a trending lane that mega-cap accounts kept hijacking (fixed by lane rules); we let aggregator funding data into an agent tool and it silently died behind a paywall (replaced with primary-source verification); and our founder-resolution once silently dropped ~2,000 founder links (found, healed, and now swept nightly). data products rot without maintenance loops - most of our engineering is maintenance loops.

where the data goes

the same signal feeds every surface: daily reports, the live feed, semantic thesis search (GET /v1/search/thesis), the public trending page, and 29 mcp tools for agents (setup guide). one pipeline, many lenses.

if you're building sourcing infrastructure yourself, the summary of everything above: curate the graph hard, keep provenance from day one, and spend most of your effort on junk. the diff is the product.

read next: how top vcs find startups before they raise · what is deal flow?