Inside Jarmin's Hiring Sprint
A startup selling autonomous AI employees to replace human engineers is hiring human engineers to build them. The paradox defines Jarmin: a five-person company promising to hand ML work to software agents races to expand the team that builds those agents.
Jarmin launched from Y Combinator's Fall 2025 batch. Its founders came from Meta's Super Intelligence Labs, Apple's machine learning division, AWS infrastructure, Lockheed Martin research, and JPMorgan Chase engineering. Co-founders Chalmers Brown and Zack Leman built LLM training infrastructure and agents for ML engineers at Meta. Shirsendu Halder shipped production ML to billions of users at Apple. Their product — an autonomous ML engineer that connects to data sources, runs experiments, and deploys production pipelines — targets the same hiring problem they now embody. "Hiring ML engineers can take 3-6+ months at $200K+ annually with costs skyrocketing for top talent," the company's YC page states.
As of July 2026, Zero G Talent's board shows six open roles at Jarmin, five already filled. The postings cluster in engineering (AI Engineer, ML Engineer, Backend Engineer, Fullstack Engineer, and Design Engineer), each carrying a $130,000 to $180,000 base plus 0.25% to 3% equity. A Founding Sales Account Executive role spans $90,000 to $300,000 with 0.25% to 2% equity, listed as San Francisco or remote. All engineering roles specify San Francisco and appeared within hours of each other, signaling a coordinated push rather than incremental backfill.
The YC launch in November 2025 announced "Jarmin.ai - You can now hire 24/7 Machine Learning Engineer employees" with a team of five. Six roles advertised simultaneously seven months later marks the inflection point where a YC demo-day company becomes a local labor-market actor. Jarmin's product automates ML engineering. Its payroll now competes for it.
The open roles ask for "any (new grads ok)" experience levels, a signal the company is casting wide rather than hunting only senior specialists. Equity bands (up to 3% for a Design Engineer) reflect pre-Series A compensation where cash conservation meets talent aggression. Every role is full-time, on-site in San Francisco. The sales role, posted a month earlier than the engineering cluster, suggests the commercial engine started before the technical build-out completed.
What This Means for SOMA's Talent Pool
The South of Market district operates at the extreme right tail of the national AI labor market. CBRE's 2025 Scoring Tech Talent report puts the Bay Area's AI-skilled workforce at 76,079 workers, up 24% from 61,497 a year earlier, yet every frontier lab and well-capitalized startup fishes from the same shallow pool. CB Insights projected the city's firms would raise $18.4 billion through 2025, representing 34% of global AI venture volume. Capital has moved faster than human capital can follow.
Jarmin's hiring push, five roles backed by Y Combinator, adds demand for precisely the talent segments with the longest search timelines and highest passive concentration. A standard backend engineering role fills in about 34 days. A GPU infrastructure or ML platform engineering role takes an average of 147 days — requiring CUDA optimization, InfiniBand networking, and distributed checkpointing at scale. That is not a premium. That is a different market entirely. Radford's Global Compensation Database and Aline Partners' AI Infrastructure Practice show the infrastructure talent pool operates at 90% passive candidate concentration with 98% employment rates among qualified professionals. Frontier model research scientists with publication records at NeurIPS or ICML face effectively zero unemployment; active application rates for posted vacancies sit below 5%.
| Role Tier | Total Compensation Range |
|---|---|
| VP Compute Infrastructure | $1.4M – $2.5M |
| VP Engineering / Distinguished Engineer (Infrastructure) | $1.2M – $2.1M |
| Staff / Senior ML Engineer | $450K – $650K |
The widely reported moderation in San Francisco technology wages (aggregate Bureau of Labor Statistics and Glassdoor data showing 3.2% annual growth, down from 8.5% in 2021) is real and completely irrelevant to frontier AI hiring. That aggregate average is pulled down by thousands of generalist software engineering roles where layoffs and oversupply have indeed moderated pay. The roles where frontier labs compete exist at the extreme right tail of the distribution. Mean compensation trends obscure this entirely.
Levels.fyi data shows the premium at every experience tier: entry-level AI engineers earn 8.6% more than non-AI counterparts; second-level AI engineers earn 11.2% more; senior AI engineers earn 10.8% more. Median AI engineer compensation peaked at $295,000 in March 2024, fell to roughly $228,500 in January 2025, then rebounded to $277,000 by March before settling in the $260,000–$269,000 range. The volatility itself signals a market where pricing is set by a handful of competing offers rather than any posted band.
Every new requisition for an ML engineer or infrastructure-adjacent role in SOMA is not competing to attract applicants. It is competing to dislodge people who are already employed, already compensated at extraordinary levels, and already receiving three to four unsolicited recruiter approaches monthly. The cost of a failed search is not just recruiter fees; it is six months of delayed model training or product development while the role sits unfilled. The incumbents are bidding up the entire market to protect their lead, and the bill lands on every startup trying to hire its first ML engineer.
How the Giants Fight Back
Google, OpenAI, Anthropic, and Meta are locked in a retention fight that makes Jarmin's five open roles look like a rounding error. The competition has moved past competing offers into territory unimaginable three years ago. OpenAI handed every technical staff member a $1.5 million bonus earlier this year, one employee reported in a public interview. A senior software engineer at the company now averages $569,000 in total compensation — $265,000 base plus $303,000 in profit participation units, the private-company equivalent of equity that could multiply far beyond public-market grants if the valuation keeps climbing. At the top end, a software engineering manager pulls $1.265 million. The floor for a technical writer sits at $144,275.
Meta, Amazon, and Google have answered with signing bonuses reaching $100 million for the most sought-after researchers, Mayer Brown's legal analysis found. Former Google HR chief Laszlo Bock called nine-figure signing packages rational, arguing the potential return on a single breakthrough model dwarfs the upfront cost. Google's own retention tactics have drawn enough scrutiny to spawn a dedicated report titled "Battle for AI Brains" that describes the standoff as unprecedented.
Money is only half the story. OpenAI's interview loop now moves faster than most Big Tech processes, a deliberate lever to close candidates before competitors can schedule a second round. Hiring managers reach out on LinkedIn before an application exists. Half the on-site interviews test collaboration and communication (walking a product manager through a system design, explaining trade-offs to a designer), not coding ability. The technical deep dive asks candidates to walk interviewers through a system they built at a previous company, evaluating whether they think in systems rather than LeetCode patterns.
Remote flexibility has become table stakes. Stripe's board data shows 46 roles added in the past seven days, several tagged US-Remote alongside San Francisco and South San Francisco listings. A senior software engineer in San Francisco carries a $224,000–$336,000 band; the same title in South San Francisco runs $190,400–$285,600. ASML added 60 roles in the same window, with a $21,000–$356,000 band and a $154,000 median. Both companies list security, infrastructure, and product counsel roles that didn't exist in their AI org charts two years ago.
The arms race has a ceiling. When OpenAI's acceptance rate drops below Ivy League levels and Google spends nine figures on a single hire, the signal is clear: the marginal value of one more elite researcher exceeds the cost of a Series A round. Jarmin's five roles in SOMA are a symptom of that distortion, not a cause.
The Ripple Effect on Real Estate and City Hall
The hiring surge at Jarmin and its peers is not just a labor-market story. It is rewriting the economics of San Francisco's commercial real estate and forcing city government to confront the same talent crunch that private companies face.
Venture capital remains the engine. In the fourth quarter of 2024, three AI funding rounds of $4 billion or more pushed quarterly capital deployed to $26.8 billion — a likely record, PitchBook data cited by Cushman & Wakefield shows. For the full year, San Francisco AI companies absorbed $48.4 billion, PitchBook's data shows. The city captured 53.2 percent of all global venture dollars flowing into the AI vertical, PitchBook's figures put. That capital converts directly into lease signatures. Tech firms accounted for 49.5 percent of new leasing activity in 2024, CBRE's data shows, and AI companies alone drove 46.4 percent of the tech sector's total, CBRE found. CBRE measured active tenant requirements at 7.9 million square feet in Q3 2025, a record, with roughly 37 percent coming from technology firms.
The office market has responded. Overall vacancy stood at 34.2 percent at year-end 2024, according to CBRE, down 20 basis points quarter-over-quarter — the first positive absorption since Q4 2019. Annual leasing volume hit 7.7 million square feet, a 28 percent jump from 2023 and the highest since the pandemic began. Several large blocks are set to hit the market in 2025, but demand is expected to absorb them. But averages mask a bifurcation. Class A trophy towers in the CBD held a 9.0 percent direct vacancy rate and commanded $109.79 per square foot, essentially flat. Meanwhile, overall asking rents slipped to $66.86, and Class A Non-CBD space fell to $67.02. BXP President Doug Linde said on a recent earnings call that AI demand is not a tower business; tenants want shorter-term, more affordable, furnished space.
Landlords are adapting. Kilroy Realty is pre-building speculative suites (fully equipped, move-in ready) before a tenant signs. The strategy landed a 93,000-square-foot lease with Harvey AI, a legal-tech firm that needed space immediately. Sublease inventory, once a deep well of discount space, has been picked through, CBRE's Colin Yasukochi said. The next wave of leasing will happen directly with landlords, not sublessors. SoMa remains the epicenter: Harvey AI at 201 Third Street, Brex at 270 Brannan. A new 42,150-square-foot lease at 201 Mission Street backfills space left by Indeed and runs through April 2028.
City government is feeling the same squeeze. The Civil Grand Jury opened an investigation into San Francisco's roadmap for identifying and implementing AI applications. A report from the City Administrator's office acknowledged that government is taking some positive steps but failing to fully embrace technology that can dramatically improve civil servant productivity and the city-public relationship. The FTC added federal pressure in January 2024, issuing orders to major AI companies examining talent acquisition practices for potential anticompetitive behavior.
The feedback loop is tight: VC funding fuels hiring, hiring drives leasing, leasing tightens prime space, and rising costs push both startups and city agencies to automate or relocate. Top-tier buildings will outperform. Mid-tier assets that have been upgraded will capture price-sensitive tenants. The market is stabilizing — but on terms set by AI companies that grow faster than lease cycles allow.
Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.