Ollama's $65M Series B Is a 37-Person Alumni Search, Not a SWE Hunt

On July 9, 2026, Ollama closed a $65M Series B led by Theory Ventures, disclosed 8.9 million monthly developers across 85% of the Fortune 500, and admitted the whole thing runs on 14 people. Founders Jeff Morgan and Michael Chiang built Kitematic, sold it to Docker in 2015, and turned it into Docker Desktop. If you are recruiting for Ollama right now and you open LinkedIn to search "Senior Go Engineer, local AI," you have already lost.

This is not a generalist SWE hire. It is an alumni-mapping problem with a very small numerator.

Why Ollama is a 37-person search, not a 66,254-person search

The right pool for Ollama is roughly 37 people who have actually shipped Docker Desktop or Kitematic, not the 66,254 Go engineers who touched llama.cpp. Ollama's engineering problem is packaging, driver plumbing, and one-command magic across four GPU backends. That is a Docker Desktop skill, not a PyTorch skill.

Here is the compression, using Refolk's index of professional profiles alongside the company's own disclosures:

The 37-person Docker Desktop alumni set has roughly 1,700 times the signal density of a naive Go+llama.cpp keyword search. Founders who do not internalize that ratio will burn a full quarter of recruiting cycles interviewing ML researchers for a job that is really about cross-compiling a Go binary that auto-detects CUDA, Metal, Vulkan, and CPU SIMD without configuration.

636,000

What the founders actually built, and what that dictates about hiring

Morgan and Chiang built a cross-platform Go binary that hides four hardware backends behind one command, which is the same product shape as Docker Desktop. That shape dictates the skill stack: Go, driver interop, installer engineering, telemetry, and a public issue tracker with tens of thousands of forks. It does not dictate transformer research.

Ollama's binary auto-detects NVIDIA CUDA, Apple Metal, cross-platform Vulkan, or CPU with SIMD, and picks a backend without user configuration. The GitHub repo has 176,000 stars, nearly 17,000 forks, and more than 67,000 community-built integrations. The engineering problem is scaling that surface, not training a new model.

The people who have done this before all sit in a few named orbits:

• Docker Desktop and Kitematic: the direct alumni, 37 profiles.
• HashiCorp: Terraform, Vagrant, Packer, all opinionated cross-platform binaries.
• GitHub Desktop, Vercel CLI, Fly.io, Tailscale, 1Password CLI: same artifact category.
• llama.cpp contributors: Georgi Gerganov's orbit, technically adjacent and pre-qualified because Ollama wraps their runtime.
• NVIDIA, Intel, AMD, Qualcomm devrel and inference-optimization teams: named distribution partners whose engineers have already integrated against Ollama's codebase.

That last bullet matters more than most recruiters realize. When a vendor's devrel engineer has spent a year filing PRs against your repo, they are not a cold candidate. They are pre-onboarded.

The alumni graph starts at two LinkedIn profiles

The canonical seed nodes for enumerating the Docker Desktop alumni pool are Michael Chiang's LinkedIn (linkedin.com/in/mchiang0610/) and Jeff Morgan's public GitHub (@jmorganca). Every other name in the 37 falls out of their first and second-degree connections plus the Kitematic and Docker Desktop GitHub commit history.

Concretely, the enumeration goes like this:

1. Start with Chiang and Morgan's LinkedIn connections filtered to "Docker" in current or past experience.
2. Cross-reference the Kitematic GitHub org's contributor graph (pre-2015) and the docker/for-mac and docker/for-win repos (post-2015).
3. Intersect with senior engineering titles (Staff, Principal, Senior Staff, Distinguished).
4. Add Solomon Hykes' and Justin Cormack's second-degree networks as expansion, since they anchor the broader Docker-founder social graph.
5. Deduplicate against current Ollama staff.

Doing this by hand takes a full week and produces a spreadsheet you will lose. Doing it inside Refolk with a plain-English query like "engineers who worked on Docker Desktop or Kitematic, senior IC titles, not currently at Ollama" collapses the same enumeration into a ranked shortlist because Refolk indexes GitHub, LinkedIn, and the open web against the same query. That is the exact gap the naive job post does not close.

The second ring: CLI-first dev tool crews

Once you exhaust the 37, the next-best pool is engineers who have shipped opinionated, cross-platform binaries at other CLI-first dev tool companies. This ring is where recruiters should look after week two, not week six.

The named companies with the right artifact shape:

• HashiCorp (Terraform, Vagrant, Packer, Consul CLI)
• Fly.io (flyctl, fly-proxy)
• Tailscale (Go daemon, cross-platform installers)
• Vercel (vercel CLI, turbo)
• GitHub (GitHub Desktop, gh CLI)
• 1Password (op CLI)
• Dagger (Solomon Hykes' current company, spiritually the closest to Docker Desktop's DNA)

The tell is not the company name, it is whether the candidate has personally owned the Windows installer, the macOS notarization pipeline, or the brew/winget/apt distribution story. Those are the three subproblems every "just works" dev tool eventually breaks on, and they are what filter Ollama's problem space from an ML engineer's problem space.

The enshittification filter nobody is running

Roughly a third of the most-qualified candidates will decline Ollama's offer over the 2025 "enshittification" backlash, and you need to know which third before you spend a recruiter loop on them. In mid-2025 a wave of blog and social posts complained that Ollama's cloud business was drawing attention away from the free project. That thread is still live in open-source purist circles.

The filter is not about excluding critics. It is about not wasting a top-of-funnel slot on a candidate who has already publicly staked a position that makes the offer a nonstarter. Pre-screen for:

• Public posts, Mastodon threads, or HN comments criticizing Ollama's cloud pivot.
• GitHub stars or contributions to explicitly "no cloud" alternatives like Jan.ai, Cortex.so, or GPT4All.
• Sponsorship of llama.cpp or maintainer status in projects with anti-monetization stances.

A candidate with any of these signals is not disqualified. They are a longer conversation, and you should route them to Morgan personally instead of a generalist recruiter.

The 37 Docker Desktop alums have 1,700 times the signal density of a Go plus llama.cpp keyword search.

Why the window closes in 18 months, not 36

Peter Fenton of Benchmark, Ollama's Series A lead and board member, has publicly predicted that open-weight models will generate the supermajority of tokens within 18 to 24 months. That is the hiring window, and it is why the Series B capital is being deployed now rather than staged over three years.

The pressure on the funnel comes from three directions at once:

1. Competitor hiring: LM Studio, Jan.ai, Cortex.so, and Nomic AI (GPT4All) are all raising or hiring against the same 37-person pool.
2. Model-lab poaching: Meta, Google DeepMind, Mistral, and MiniMax are named Ollama distribution partners; their local-inference teams are the exact profile Ollama needs, and those teams also have retention packages.
3. GPU-vendor absorption: NVIDIA, Intel, AMD, and Qualcomm are quietly building out their own local-runtime devrel orgs and hiring the same people.

If Ollama wants to grow from 14 to, say, 40 engineers over 18 months, that is roughly 1.5 hires per month against a pool the entire industry is competing for. The math only works if the sourcing funnel is depth-first, not breadth-first.

The 636,000 ratio is a hiring constraint

Every Ollama employee currently supports 636,000 monthly developers, which means each new hire changes the culture by roughly 7% and each mishire is a compounding problem for a year. This is not a bragging stat. It is a hiring constraint that rules out the Anthropic-style Series B binge.

Practically, it means:

• Hiring in single digits per quarter, not per week.
• No "we'll figure out the org design later." The first 10 post-Series B hires are the org design.
• Referrals from the 14 will outperform any inbound channel by an order of magnitude, because the existing team knows exactly which of the 37 they want to work with again.
• External sourcing has to be surgical. The plain-English query "shipped Docker Desktop, now at a company they are quietly unhappy at" is the exact shape of work Refolk was built for, because it collapses signal from LinkedIn, GitHub, and the open web into one ranked list instead of three tabs.

0.056%

What Y Combinator W21 quietly gives Ollama

Ollama is a W21 YC company, and Morgan and Chiang previously ran a dev-infra startup called Infra, which widens the warm-intro layer beyond Docker. The W21 batch plus the Infra alumni graph is the second seed set most recruiters skip.

Concretely, the warm-intro layers are:

• YC W21 batchmates: co-founders and early engineers who overlap with Morgan and Chiang from batch programming.
• Infra alumni: small cohort, high loyalty, already worked with the founders.
• Docker Desktop team leads pre-2020: worked with Morgan and Chiang directly.
• Kitematic contributors pre-2015: the smallest and warmest ring.

You can run this enumeration inside Refolk by asking for "people who overlapped with Jeff Morgan or Michael Chiang at Docker, Infra, or YC W21." The point is not the tool, it is that the four rings above are the entire warm funnel, and the recruiter who maps them in week one is the recruiter who fills roles in month three.

FAQ

How many Docker Desktop alumni are actually reachable for Ollama?

Roughly 37 profiles surface when filtering strictly on Docker Desktop or Kitematic product work with senior engineering titles, per Refolk's index. Of those, a meaningful fraction are already at Ollama, at Dagger with Solomon Hykes, or at Docker itself on retention packages. The realistically reachable pool is closer to 20, which is why the second ring of CLI-first dev tool companies matters so much and why depth-first sourcing beats breadth-first sourcing here.

Should Ollama hire ML researchers or infrastructure engineers?

Infrastructure engineers, almost exclusively. Ollama's binary is a Go program that auto-detects CUDA, Metal, Vulkan, or CPU SIMD and picks a backend without configuration. That is a packaging, installer, and driver-interop problem. The model work happens upstream at Meta, Google DeepMind, Mistral, and llama.cpp. Hiring PyTorch researchers for a Go-plus-drivers job is the single most common mistake founders make when they read "AI company" and skip the actual product surface.

What is the fastest way to enumerate the Docker Desktop alumni pool?

Start at Michael Chiang's LinkedIn and Jeff Morgan's GitHub, walk the first and second-degree connections filtered to Docker experience and senior IC titles, cross-reference the Kitematic and docker/for-mac contributor graphs, and deduplicate against current Ollama staff. A plain-English query in Refolk does the same enumeration across GitHub, LinkedIn, and the open web in one pass, which is the specific friction the two-founder-plus-spreadsheet approach hits by day three.

Why does the hiring window close in 18 months?

Because Benchmark's Peter Fenton has publicly said open-weight models will generate the supermajority of tokens within 18 to 24 months, which means the local-runtime category consolidates in that window. LM Studio, Jan.ai, Cortex.so, and Nomic AI are all hiring against the same 37-person Docker Desktop pool, and NVIDIA, Intel, AMD, and Qualcomm are absorbing local-inference engineers into their own devrel orgs. Ollama's Series B capital is being deployed now for exactly this reason.