# Tracing a Founder Factory's Alumni to Their Next Startups

*You can start from one founder factory and finish with a ranked list of pre-round alumni startups, each founder resolved to a reachable identity.*

- Canonical URL: https://www.refolk.ai/guides/founder-factory-alumni-startups
- Pillar: Investing and deal sourcing
- Format: Teardown
- Published: 2026-09-18
- Last reviewed: 2026-09-18
- Reading time: 16 min
- Keywords: founder factory deal sourcing, how to track startup mafia alumni, find founders spinning out of a company, trace unicorn alumni new startups, source pre-round founders from alumni network, map where a company's employees go next

## Key takeaways

- Roughly five fundable spinouts per unicorn is the sanity-check number: 203 European and Israeli unicorns produced 1,018 alumni startups, so a filter returning wildly more or fewer is miscalibrated.
- In Refolk's index, the substring 'Stripe' plus a founder or CEO title returns 497 US profiles versus 121 UK ones, but that raw count also catches 'Stride Labs' and 'Strike', so exact-employer resolution must precede ranking.
- A founding-engineer job post at a company with no website is the strongest cheap formation signal, because it proves a team is forming rather than one person tinkering.
- Founders with three or more years of prior work experience are 85% more likely to launch a successful startup, so tenure and seniority filters beat a hunt for prodigies.
- The best factory is a medium win, not a mega-win: a $300M to $2B exit that endured near-death produces more enduring founders than a decacorn whose alumni are distracted by wealth.
- 56% of ex-unicorn companies were founded in the same city as the unicorn, so geo-filtering to the factory's home city cuts noise cheaply before any scoring.

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.

| Criterion | What qualifies | Why it matters |
| --- | --- | --- |
| Exit size | Roughly $300M to $2B | Life-changing wealth kills founding motivation |
| Adversity | Went through near-death trying times | Survivors learn to build under pressure |
| Endurance | Held together post-exit | Networks and trust persist into new ventures |

**~5 - Fundable spinouts per unicorn (sanity-check baseline)**

203 European and Israeli unicorns produced 1,018 alumni startups, which divides to about five each.

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.

> **Note:** The worked example
>
> For this teardown I run the search as if the factory were Stripe, because the dossier carries real counts for it. Stripe is arguably too successful under the medium-win rule, so treat its numbers as an upper bound and a calibration source, not the ideal target. The method is identical for a $500M fintech exit; only the counts shrink.

## 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.

| Source | What it proves | What it misses |
| --- | --- | --- |
| Professional-profile fields | Tenure, title, that someone left a job | That they are founding, not on sabbatical |
| Incorporation registries | A legal entity exists, with named officers | Founders who have not filed yet |
| Domain registration | Intent to build something | Everything; too noisy alone |
| GitHub org and new repos | Code is being written | Whether it is a company or a hobby |
| SEC Form D, SH01 filings | Money has moved, sometimes pre-press | Bootstrapped or unfunded ventures |

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.

**85% - More likely to launch a successful startup**

Founders with three or more years of prior work experience, per the Azoulay/Kim/Miranda research.

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.

| Company | Alumni who became founders | Context |
| --- | --- | --- |
| Bain & Company | 8.13% | Highest US rate |
| Goldman Sachs | 5.92% | Most future founders in finance |
| Twitter | 6.17% | High per-capita for its size |

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.

| City | Unicorns | Alumni startups | Same-city % |
| --- | --- | --- | --- |
| Tel Aviv | 27 | 108 | 81% |
| Paris | 22 | 125 | 75% |
| Berlin | 24 | 138 | 70% |
| London | 27 | 168 | 69% |
| Stockholm | 11 | 98 | 59% |

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.

#### Stripe worked example, from raw pull to pre-round list

| Stage | Figure | Note |
| --- | --- | --- |
| Raw substring pull (US, index) | 497 | "Stripe" + founder/CEO title, before dedup |
| Exact-employer resolved | 125 | matches Dealroom's alumni-startup count |
| Tenure + geo filtered | 40 | early hires, SF/London, 3+ years |
| Two-plus formation signals | 12 | incorporation, code, or founding-engineer hire |
| Round-ready ranked | 8 | scored and deduplicated |

*Each filter discards most of the prior stage, which is expected when isolating a rare founding tail.*

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.

> **Rule:** Two signals, one window, or it does not count
>
> Tag a candidate "forming" only when two or more independent formation signals co-occur inside roughly a two-week window. A single departure, a single domain, or a single repo is a watch item, not a lead. This rule is what keeps hobby projects and sabbaticals out of the deal list.

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.

#### From one factory to a ranked pre-round list

1. **Pick the factory and justify it** - Choose a company with a successful but not crazy-successful outcome, roughly a $300M to $2B exit, that survived near-death and endured post-exit. Write a one-line thesis for why this factory produces founders now.
2. **Pull the raw alumni set** - Query professional-profile data and any curated mafia page for former employees, capturing tenure dates and last title. Run both a curated seed and an unfiltered pull, because the curated list under-counts stealth and the raw pull over-counts.
3. **Filter by tenure and seniority** - Keep early hires and operators with three or more years of experience, since that cohort founds at higher rates. The output is a shortlist ranked by founding propensity.
4. **Detect formation signals** - For each shortlisted alum, check incorporation records, domain registration, GitHub org, SEC Form D, and founding-engineer job posts. Require two or more signals inside one two-week window, then tag each candidate forming or not-forming.
5. **Resolve identity and co-founders** - Match each name to one reachable person using attributes beyond the name, such as tenure dates, location, and prior title, then inspect incorporation officers and early hires for a second founder. Each venture ends with every founder resolved to a single profile plus a contact path.
6. **Score round-readiness and rank** - Score each surviving venture on signal-stack strength, thesis fit, and evidence of a forming team. The output is a ranked, deduplicated pre-round list.
7. **Reach out** - Give one legitimate reason for reaching out, explain the thesis connection, and acknowledge the work may be early. Do not list every signal or imply access to private information, then log first replies.

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.

I ran this search: `Former Stripe employees, joined before 2018, now listed as founder or CEO of a company under two years old, based in San Francisco or London.` - [see the full result list](https://www.refolk.ai/s/y16ykv0s8w).

*Returns early-hire alumni already carrying a founder title, pre-filtered on tenure and geography so you start at step four rather than step two.*

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.

#### Name to reachable founder

1. **Candidate name** - One alum flagged forming, name only
2. **Attribute match** - Confirm on tenure dates, location, prior title
3. **Merge records** - Combine profile, incorporation, repo above threshold
4. **Co-founder sweep** - Read incorporation officers and early hires for a second identity
5. **Contact path** - Resolve each founder to a reachable channel

*Resolution moves from a shared name to a single confirmed identity plus every co-founder on the cap table.*

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.

> **Watch out:** String matching inflates every count you never audit
>
> The substring "Stripe" in Refolk's index also captured "Stride Labs," "Strike," "Stripe Partners," and "Green Stripe Media," so the raw 497 US figure is an upper bound. Every downstream number you compute on an unaudited pull carries this contamination. Require the exact former-employer entity before you count anything.

- **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.

| Proxy | What it proves | When it lies |
| --- | --- | --- |
| Signal-stack strength | Two-plus signals in a window means real formation | A single loud signal can be theater |
| Thesis fit | Founder experience aligns with your fund's focus | A resume match without a formed team is empty |
| Team-formation evidence | A technical co-founder or first hire exists | A solo incorporation may stall for months |

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.

**Round-readiness score (0 to 3 per axis, rank on the sum)**

```
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
```

*Adjust the axis weights to your stage. Anything scoring 0 on team formation goes to the watchlist, not the deal list.*

Before you call the job done, run this check.

#### Before this list goes to the partnership

- [ ] Every count is on exact-employer matches, not a substring pull
- [ ] Each candidate carries two or more formation signals inside one window
- [ ] Each founder is resolved on tenure, location, and prior title, not name alone
- [ ] Every venture has been swept for a co-founder and second identity
- [ ] The list is deduplicated, so one venture appears once with all founders attached
- [ ] The total spinout count is near the ~5-per-comparable-unicorn baseline
- [ ] Each round-ready lead has a resolved contact path

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.

## Frequently asked questions

### How many spinouts should one founder factory realistically produce?

Use roughly five fundable spinouts per comparable unicorn as your sanity-check number. Dealroom counted 1,018 alumni startups from 203 European and Israeli unicorns, which divides to about five each. Stripe is an outlier at 125 alumni startups because it is far larger. If your filter returns wildly more or fewer than five per comparable factory, it is miscalibrated and you are either counting side projects or missing stealth founders.

### What single signal best separates a real startup from a side project?

A founding-engineer job posting at a company with no website is the strongest single formation signal. It proves a team is forming and someone is spending to hire, which a domain registration or a lone GitHub repo does not. In isolation a domain produces too many false positives to trust. The predictive pattern is multiple signals about one person inside a two-week window: incorporation plus a matching domain plus new code, for example.

### Which company makes the best founder factory to hunt?

A company with a successful but not crazy-successful outcome, roughly a $300M to $2B exit, that went through near-death and endured post-exit. When a company is too successful, would-be founders get distracted by life-changing wealth and never build again. That is why mega-wins like the largest platforms can be weaker hunting grounds per capita than a medium exit whose alumni still have something to prove.

### How do I avoid counting the wrong people when searching by former employer?

Require the exact former-employer entity, not a substring, before you count anything. In Refolk's index, matching on 'Stripe' also captured 'Stride Labs', 'Strike', and 'Stripe Partners', so the raw 497 US count is an upper bound. Confirm each alum on tenure dates, location, and prior title before ranking, and check incorporation officers so you do not collapse two different people with the same name into one.

### Does leaving a company mean someone is founding?

No. A quiet departure without an announced next step is only a signal, and it is often a sabbatical or a lateral move. Pair every departure with a formation signal such as an incorporation, a founding-engineer hire, or new code inside a tight window. Two people from the same team leaving within weeks strengthens the read, but on its own an employment change never confirms a company exists.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/guides/founder-factory-alumni-startups*