Profiles name the field, rarely the tools
50% of US machine learning engineer postings ask for PyTorch. 7.3% of the 3,161 machine learning engineers in San Francisco list it. A search that filters on a tool keeps only the people who wrote it down.
On this page
What we counted
Profiles visible through the Metix AI Platform on September 30, 2026 whose current title matches "machine learning engineer" and whose location is San Francisco. That is 3,161 people. Then, for each of three tools, how many of those 3,161 name the tool in the skills section of their profile, and how many open US postings for the same title name it in their description.
Every count here is one API call, and all of them are reproduced by the find engineers by skill and city use case, which publishes the queries and what they cost.
The field, not the tools
7.3%of San Francisco machine learning engineers name PyTorch among their skills: 230 of 3,161.
Skills lists are not empty. Every one of the 3,161 lists at least one skill. What they list is the field. Tool names are the exception: Python, the language nearly all of this work is written in, appears on 40% of the profiles. TensorFlow appears on 7.8% and PyTorch on 7.3%.
Source: Casebook use case, skills table
Job postings say the opposite. Of the 1,658 open US postings for the same title, 50% name PyTorch in the description and 35% name TensorFlow. Employers ask for PyTorch far more often than candidates write it down, and TensorFlow, the older framework, still appears on more profiles than PyTorch does. One reading is that a skills list is written once and rarely updated, so it records the tools that were current when it was written.
Why a tool filter undercounts
A search that requires skills match "pytorch" does not find people who use PyTorch. It finds people who wrote PyTorch in one particular section of their profile. In San Francisco that turns 3,161 candidates into 230. The 2,931 it drops are people who did not list PyTorch. With half of the postings for the role asking for it, many of them likely use it; the counts cannot say how many, and that is the problem: the filter decides for you without showing what it removed.
The same filter is the last step in the use case's ledger, where it keeps 7.3% of the row above, the smallest share of any step. The country keeps 33% and the city 14%. A city is meant to cut hard, since most engineers live somewhere else. When a skill filter cuts twice as hard as a city, it is usually measuring how people write their profiles, not the market.
What to search on instead
- Filter on what people state about themselves. Title, location and seniority are filled in on nearly every profile. Use them to define the pool.
- Use the tool to rank, not to exclude. Read the pool's records and put the people who name the tool first, instead of dropping everyone who does not.
- Look where people describe work. A headline or a job title sometimes names the tool. In San Francisco 41 of the 3,161 name PyTorch in their headline, against 230 in their skills, so it widens little here; check it before relying on it.
- Count each step first. A count costs 1 API Credit. If one filter keeps far less than the others, you see it before you pay to read records.
Limits
One city, one title and one day. Counts are profiles visible through the Metix AI Platform, not everyone in the market. A skills match finds the word in the skills section and does not say how well the person knows the tool. The reading about when skills lists were written is an interpretation, not something the counts measure.