The search, exactly as it was typed
No query operators, no filter form. This is the whole input:
StarHunt read that sentence, mapped it to the technologies and role it indexes, filtered to the San Francisco metro area, and ranked what it found by evidence of real work. What follows is the top of that result list.
The five profiles it ranked first
Every name links to a public GitHub profile, so you can check the claim yourself. That is the point of sourcing on public work: nothing here needs to be taken on trust.
Founding AI engineer at Prime Intellect, MIT computer science before that. He built proton-drive-sync, his own client sitting at 669 stars, and he ships to opencode. The profile reads like an infrastructure engineer who moved into AI products rather than the reverse, which is the harder direction and usually the more useful one.
Contributor to anomalyco/opencode. Own most-starred project: DamianB-BitFlipper/proton-drive-sync, 669 stars.
- 69 followers
- 61 public repos
- Active 2 weeks ago
- AI Product
- Software Eng
- llm
- python
- typescript
Chief Architect at OpenClaw, previously Qantas, Airbyte and Microsoft, and a lecturer at MIT. He contributes to LiteLLM, which is the routing layer a large share of production LLM applications actually run on. Someone who has both operated at airline scale and works on the plumbing everyone else depends on is a rare combination for an AI product role.
Contributor to BerriAI/litellm. Own most-starred project: vincentkoc/tokenjuice, 503 stars.
- 2,124 followers
- 37 public repos
- Active 2 weeks ago
- AI Product
- Infra / DevOps
- agents
- huggingface
- llm
VP of AI at Elastic, and before that founder and CEO of Jina AI, which Elastic acquired. What makes the profile interesting is that he never stopped writing code: he contributes to llama.cpp and his most-starred recent project is claudecode-telegram, a Telegram bridge for Claude Code, at 607 stars. An executive with an exit behind him who still ships inference code on the weekend is not someone a resume search finds.
Contributor to ggml-org/llama.cpp. Own most-starred project: hanxiao/claudecode-telegram, 607 stars.
- 4,219 followers
- 61 public repos
- Active 2 weeks ago
- AI Research
- AI Product
- inference
- agents
- llm
Building AI for tax professionals at Additive, now part of Thomson Reuters. Three-time founder, and at Salesforce before that through two acquisitions. He wrote the MCP server for the Zotero API, which is exactly the kind of unglamorous integration work that separates people who ship AI products from people who demo them.
Own most-starred project: kujenga/zotero-mcp, 160 stars.
- 54 followers
- 49 public repos
- Active 3 weeks ago
- AI Product
- AI Research
- agents
- jax
- python
A four-word bio, "Containers. Linux. Container Linux.", and a career to match: currently at Red Hat, with deep low-level systems expertise. He contributes to graylog2-server, and his own work is testing and infrastructure tooling rather than anything with an audience. His most-starred repository sits at 28 stars, which is the whole point: stars measure attention, and infrastructure work is the least watched kind. On any keyword search he is invisible, because he does not describe himself at all.
Contributor to Graylog2/graylog2-server. Own most-starred project: nstielau/cucumber-varnishtest, 28 stars, a testing harness for Varnish.
- 76 followers
- 49 public repos
- Active 2 months ago
- AI Product
- Software Eng
- rag
- agents
- python
Profiles shown with public GitHub data only. Anyone can remove themselves from the index in one click.
What a keyword search would have missed
The interesting part of this list is not who is on it, it is that a normal search would never have reached most of them. Three specifics from the five profiles above, all checkable in a browser.
Three of the five have under 80 followers. The spread runs from 54 to 4,219, a factor of eighty. GitHub's own user search ranks close to follower count, so Aaron Taylor at 54 and Damian Barabonkov at 69 sit far below the fold on any query that would surface Han Xiao at 4,219. Ranking on merged contributions instead puts them side by side, which is where they belong: influence and building are different things, and only one of them is the job.
Two of the five never say what they do. Nick Stielau's entire bio is "Containers. Linux. Container Linux." Damian Barabonkov's is a company handle and "prev @mit cs". Neither contains the words "AI", "LLM", "machine learning" or "Python". Search python AI location:"San Francisco" and neither exists. Their work says it instead: RAG and agent projects for one, an inference company and a 669-star sync client for the other.
One of the five is not in San Francisco. Han Xiao's GitHub location reads Mountain View, CA. A literal location:"San Francisco" filter drops him. StarHunt normalizes every location to its metro area first, so Mountain View, Palo Alto and Sunnyvale all resolve to the San Francisco hub, the same way Massy and Versailles resolve to Paris. Nobody searching for a Bay Area hire means to exclude Mountain View.
None of this requires trusting the tool. Every claim above is a link away.
Run the same search on your own role
This took one sentence and under 30 seconds. Free accounts get three searches, with full profiles, contact details and unlimited exports, and no credit card.
Start for free →Curious how the ranking works? Read the methodology, or see the wider picture in the GitHub sourcing guide.
Sourcing AI engineers, answered
How do you find AI product engineers in San Francisco?
Start from the repositories that define the AI product stack, inference servers, agent frameworks, RAG tooling, model serving, and look at who actually merges code into them. Filter that pool to the San Francisco metro area, then rank by contributions over the last twelve months and by the languages in each person's own repositories. That surfaces people building AI products today, rather than people who list machine learning as a skill.
What is the difference between an AI product engineer and an ML researcher?
An AI product engineer ships user-facing systems built on models: retrieval, agents, evaluation, latency, cost, and the application around them. A researcher advances the models themselves. The GitHub evidence differs accordingly: product engineers contribute to inference, orchestration and application repositories, researchers to training and modelling code. StarHunt indexes both and separates them by the projects a person actually contributes to.
Why does GitHub search miss these engineers?
GitHub user search matches keywords in a bio and ranks close to follower count. Someone who merges into a major inference or agent framework every week but writes "software engineer" in their bio never appears. The work is public, the bio is not searchable evidence of it.
Are these developers looking for a job?
Most are not, which is the point. Sourcing from public work reaches people who are not on job boards. StarHunt shows who has marked themselves open to work on GitHub, but never limits results to them.