Tracing a Founder Factory's Alumni to Their Next Startups

This teardown follows one founder factory end to end into a deal list. It is for early-stage investors, platform and talent partners, and angels who want to reach alumni-founded startups before they announce. By the end you will have a ranked, deduplicated list of newly formed companies started by a single company's alumni, each founder resolved to a reachable identity and each venture scored for round-readiness. I carry one example the whole way through, with the real intermediate counts, the forks, and the wrong turns.

Two existing patterns already cover pieces of this. A stealth-founder watchlist tracks one person. A talent-flow map ranks where people go in aggregate. Neither takes a single alumni network and finishes with an investor's deal list. That is the gap this fills.

What a founder factory is, and why the medium win beats the mega win

A founder factory is a company whose former employees start an outsized number of new ventures. The counterintuitive part: the best factory is a medium win, not the largest exit you can name. The dossier is blunt about the mechanism. When a company is too successful, people who would have become part of the mafia get distracted by their wealth, so a $300M to $2B exit that went through near-death and endured post-exit is a better hunting ground than a decacorn.

That is why per-capita matters more than volume when you pick a factory. The largest platforms produce the most founders by raw count simply because they employ the most people, but normalized pipelines like DeepMind and Palantir surface as the real high-yield engines. Volume tracks headcount. Alumni-years track propensity.

Before committing to a factory, screen it against three criteria drawn from the research.

Anchor your expectations to that number. Stripe is an outlier at 125 alumni startups worth $1.9T because it is far larger than a typical unicorn; a normal comparable factory should yield close to five fundable spinouts. If your pipeline returns fifty or two, the filter is broken before you score anything.

Which public sources surface alumni and their founding status

Five public sources surface a company's former employees and whether they are founding, and each misses something the others catch. No single source is dispositive, which is why the whole method is signal-stacking rather than one clean query.

The employment signal is the weakest on its own. A title shift to "Building something new," "Stealth," or a blank current position tells you someone left a job, not that they are founding. A network signal adds that they are recruiting people to join them, which is a different and stronger read.

Registries are the cleanest yes/no. In the US, the Delaware Division of Corporations lets you search by officer name and incorporation date. In the UK, Companies House publishes SH01 share-allotment forms that can reveal stealth funding before any press coverage. The four places stealth companies leak, per the dossier, are certificate transparency logs, professional-profile bio changes, indexed-but-unlinked Notion or Linear pages, and SEC Form D filings.

Domain registration is the trap. GDPR has made individual identification from WHOIS harder in privacy-protective jurisdictions, and in isolation domain registrations produce too many false positives to be a primary signal. Keep it as corroboration, never as a trigger.

Which alumni are worth filtering to, and by what

Filter to experienced operators, not prodigies. The founder-genius myth does not survive the data. Among TechCrunch-award founders the average age at founding was 31, and for Inc.'s 2015 fastest-growing startups it was 29. The clean signal is prior experience.

So the first hard filter is three or more years of experience before founding. The second is hire number: early hires found at higher rates than late joiners, because they saw the company scale and absorbed the playbook. The exact founding rate for the specific second-to-tenth-hire band is not established publicly, so treat hire-number as a ranking input, not a hard cutoff. What the research does establish is company-wide alumni founding rates.

Use these as base rates, not targets. Even at the top of the table, founding is rare: more than 91% of Bain alumni never found anything. Your filter is trying to isolate a small tail, so expect to discard most of the raw pull.

Geography cuts noise cheaply. 56% of companies founded by ex-employees were founded in the same city as the unicorn, because co-founder networks and investor ties are geographically sticky.

Geo-filtering to the factory's home city keeps most of the diaspora and drops the long tail. In my worked example, I hold San Francisco and London, since Stripe's population concentrates there.

Detecting formation: the two-week window

The single test that separates a formed startup from a side project is signal-stacking inside a tight window. A person who incorporates, registers a matching domain, and starts pushing code to new repositories within the same two-week period exhibits a highly predictive pattern. One signal is noise. Two or more, co-occurring, is a company.

Those middle counts are illustrative of the shape, anchored to the two figures the dossier does establish: 497 raw US matches and 125 real alumni startups. The lesson is the drop rate. Roughly three quarters of the raw pull evaporates at exact-employer resolution alone, and most of what survives is not forming a company right now.

The strongest single cheap signal is a hire, not a filing. A founding-engineer job post at a company with no website is among the strongest formation signals available, because even a stealth company in its earliest weeks needs at least one technical co-founder or first engineering hire. A landing page proves nothing; a job requisition proves someone is spending money to build a team.

Real-time monitoring of new incorporations also filters out shell companies and holding structures, so you surface real founders rather than a lawyer's registered-agent boilerplate. If an incorporation has no matching domain, no code, and no hire behind it, hold it on a watchlist and re-check in a month rather than promoting it.

The procedure, end to end

Here is the full method. Each step names who does it and roughly how long it takes, so you can staff it.

The fork that matters most is at step two. Seed from a curated mafia page and you get clean names but miss the stealth founders who have not announced, which are the ones worth reaching first. Run only a raw pull and you drown in string-match contamination. The answer is to run both and merge, treating the curated list as a floor and the raw pull as the search space.

The hardest step in practice is identity resolution, so I gave it its own section below. Once you can turn a name into one reachable person plus their co-founder, the query itself is a single ask.

Where that removes friction is the raw pull and the first two filters at once. Instead of exporting an alumni list and hand-filtering tenure and geography, Refolk takes the constraint in plain English and returns the shortlist. You still owe the formation-signal check and the identity resolution yourself, because those depend on registries and judgement no single query settles.

Resolving one name to one reachable founder, plus the co-founder

Match-then-merge, using attributes beyond the name, then look for the second founder. A name match alone is not dispositive of identity because many names are shared, so resolution depends on confirming tenure dates, location, and prior title before you treat two records as one person.

The formal process has two moves. The matching step calculates pairwise record similarities on attributes like name, and the merging step combines records above a similarity threshold. In practice that means: given "Alex Chen, ex-Stripe, SF," you confirm the incorporation officer, the professional profile, and the GitHub account are the same Alex Chen by cross-checking tenure and location, not by trusting the name.

The co-founder sweep is the step people skip, and it is where you lose the real builder. If you resolve only the loudest name you may drop the technical co-founder who is doing the work. Inspect incorporation officers and the earliest hires for a second person, then resolve that identity the same way. The inverse of the name-collision logic is useful here: if two people in a network have very similar names and closely related work they are probably the same person, so two distinct names sharing one venture and one filing are almost certainly genuine co-founders.

How this goes wrong: failure modes and false positives

This is the section that earns the guide. Every filter above has a characteristic way it lies, and most of them inflate your count silently. Here is what each looks like when it fails and how to check it locally.

• String-match contamination. A keyword pulls unrelated firms whose names contain it. Check: require the exact former-employer entity, not a substring, before ranking. This is why 497 raw matches resolve to 125 real alumni startups.
• Left is not founding. A quiet departure without an announced next step is only a signal, and it is often a sabbatical or a lateral move. Check: pair every departure with a formation signal; two people from the same team leaving within weeks strengthens the read but never confirms it alone.
• Domain-only false positive. In isolation, domain registrations produce too many false positives to be reliable as a primary signal. Check: require a co-occurring incorporation or code-repo signal.
• Hobby project versus company. A GitHub org or landing page alone can be a side project. Check: look for a founding-engineer hire or an incorporation, not just infrastructure.
• Name collision. Two different people share a name and a false positive collapses them. A match on only the name is not dispositive because many names are shared. Check: confirm on tenure dates, location, and prior title.
• Missing the co-founder. Resolving only the loudest name drops the technical co-founder who is the real builder. Check: read incorporation officers and early hires for a second resolved identity.
• Curated-list under-count. Seeding only from a published mafia list misses stealth founders who have not announced. Check: run a raw alumni pull in parallel.
• Too-successful factory. Companies too successful may not create enduring mafias because wealth distracts would-be founders. Check: apply the exit-size and near-death criteria before committing.

Every filter in this method inflates your count silently until you audit exactly which entity it matched.

Two of these deserve extra weight because they are asymmetric. String-match contamination and curated-list under-count push in opposite directions: one over-counts noise, the other under-counts the very stealth founders you most want. Running the exact-employer audit and the parallel raw pull together is the only way to catch both at once.

Scoring round-readiness and staying current

No canonical public rubric for earliest-stage round-readiness exists; treat that as an honest limit and score on defensible proxies. The dossier is explicit that this is not established publicly, so anyone selling a precise round-readiness number is overclaiming. What holds up are three primary-source proxies.

The strongest cases involve founders whose experience aligns with the fund's thesis and who have recently changed direction, and experienced operators often leave a role months before announcing. Team formation is the gating signal: even a stealth company in its earliest weeks needs at least one technical co-founder or first engineering hire, so a venture with no second person scores lower regardless of pedigree.

Founder: __________  Venture: __________
Signal-stack strength (0-3): __ (0=one signal, 3=three-plus in one window)
Thesis fit (0-3):           __ (0=off-thesis, 3=direct)
Team-formation evidence (0-3): __ (0=solo, 3=technical co-founder + first hire)
Total (0-9): __   Verdict: reach now / watch / drop

Before you call the job done, run this check.

Keep the work current by treating formation signals as perishable. The two-week window that makes a signal-stack predictive also means a lead a quarter old has either raised or stalled, so re-run the raw pull and the incorporation sweep on a monthly cadence and promote watchlist items only when a second signal lands. When you reach out, give one legitimate reason, explain the thesis connection, acknowledge the work may be early, and do not list every signal or imply access to private information. That last discipline is what keeps a diaspora warm across the many ventures you will not fund this time.