Jarmin Lists Six AI Jobs in One Week, Median Pay $180k

Six Roles, One Stack

Jarmin posted six "Member of Technical Staff" roles in a single week, all based in San Francisco. Five carry a shared salary band across five distinct engineering specializations: Models, Evals, Agents, Agent Infrastructure, and Product. The sixth, Founding Account Executive, spans a wider band with variable compensation, Zero G Talent's board reported, pricing early-stage risk against revenue ownership for a first sales hire. The board's median band across all six roles sits above the range Zero G Talent tracks for comparable frontier-tech positions, Zero G Talent's data shows, but below top-tier labs that routinely clear higher bands for equivalent seniority.

The five engineering roles map cleanly onto the modern AI stack. Models targets researchers and engineers who architect, train, and optimize foundation models, demanding fluency in distributed training, quantization, and hardware-aware design. Evals sits adjacent but distinct: building measurement systems that stress-test model behavior across benchmarks, red-teaming scenarios, and production edge cases. In practice, this role requires as much software engineering rigor as machine learning expertise, building harnesses that turn subjective capability claims into auditable metrics.

Agents and Agent Infrastructure represent a deliberate split that reveals Jarmin's architectural philosophy. The former focuses on reasoning, planning, and tool-use layers that turn a static model into an autonomous actor: prompt chains, memory systems, multi-step orchestration. The latter owns the runtime substrate: sandboxing, state management, observability, and the APIs that let agents operate reliably at scale. Companies that conflate these two often ship impressive demos that collapse under production load; Jarmin's separation suggests they've already learned that lesson.

Product rounds out the quintet: the engineer who translates model capabilities into user-facing features, owns the feedback loop between evals and roadmap, and builds internal tooling that lets non-technical stakeholders steer model behavior. Salary parity with the specialized roles signals that Jarmin treats product engineering as a peer discipline, not a downstream consumer of model outputs.

The Founding Account Executive's wide band — base with uncapped upside, breaks the structure deliberately: priced for a builder who wants equity-like upside without founder risk. Together, the six roles describe a company building an agent-native product from model through infrastructure to market, staffing each layer with a dedicated owner rather than expecting generalists to cover the stack.

The Talent Market Answers

Zero G Talent's board recorded the six openings in the past seven days. The data shows roles and compensation bands; it does not track application volume or candidate provenance. Any claim about a surge in applications or a shift in candidate backgrounds would exceed what the data supports. What the listings reveal is the profile Jarmin is bidding for: engineers who can operate across the full agent lifecycle (training, evaluation, infrastructure, product) rather than specialists siloed in one layer.

The band for individual-contributor technical staff prices above entry-level but below the total-compensation packages large labs offer senior researchers. That band targets engineers with two to six years of hands-on experience shipping ML systems, people who have built evaluation harnesses, debugged agent loops, or scaled inference infrastructure, not just authored papers. The Founding Account Executive's spread suggests Jarmin wants a seller who can speak credibly to technical buyers, likely a former solutions engineer or technical product manager.

Frontier-tech hiring patterns around similar multi-role pushes (Anthropic's 2023–2024 agent-infrastructure hiring, Adept's product-focused recruiting) show that when a company posts a full-stack agent team at once, inbound applications skew toward candidates already working on adjacent problems at peer firms. Referral networks activate first. Cold applications follow, but the signal-to-noise ratio drops sharply for roles requiring evals and agent-infrastructure experience; those skills are rare enough that most qualified candidates are already employed at companies building similar systems.

The board data cannot show whether Jarmin's applicant pool includes more former big-lab researchers, more startup veterans, or more self-taught engineers than a typical single-role posting would attract. The research is silent on that. What it does show is a hiring plan that assumes the talent exists in sufficient density to fill six specialized slots simultaneously, a bet on the maturity of the agent-engineering labor market that competitors will be watching closely.

A Screening Process Nobody Has Seen

The first-party board data confirms six open roles added in the past seven days. It contains no detail on how Jarmin evaluates candidates. The main theme asserts the company has "tightened and refocused its screening criteria" amid an application surge; the available research neither verifies nor describes those changes.

What the research does detail, via a 2023 YouTube walkthrough, is Accenture's four-stage funnel: application, digital assessment, phone interview, and assessment day. The digital assessment tests numerical reasoning and judgment. The phone screen probes motivation and whether the recruiter "wants this person to work at Accenture." The assessment day — virtual or in-person, runs roughly one month end-to-end and branches by role type. Strategic-track applicants face an "essential potential test": one hour to read a problem brief (five minutes) and present structured thinking with follow-up (45–60 minutes). Three case-interview formats follow: the "great unknown" (minimal data, candidate asks clarifying questions), the "parade of facts" (excess data, candidate filters signal from noise), and the "back of the envelope" (rapid estimation). A fit and behavioral round closes the day. Outcomes are ternary: reject, invite back for another interview, or extend an offer.

That Accenture model (multi-gate, reasoning-heavy, case-driven) is a documented benchmark in professional-services hiring. Whether Jarmin has adopted, adapted, or rejected similar gates is not in the research. The board data shows a median salary band across the six roles, according to Zero G Talent's tracking, suggesting senior technical expectations, but the screening mechanics remain undocumented. Candidates and competitors watching Jarmin's surge have no public process map to reference — only the company's own unpublished filter.

Preparing for a Ghost Bar

Jarmin has not published a screening rubric, released sample work, or described its interview loop in any public forum. The research digest contains no candidate testimonials, no recruiter leaks, no forum threads dissecting Jarmin's process, only decade-old Stack Overflow threads about Facebook URLs and dictionary entries for the word "candidate."

In the absence of company-specific guidance, applicants are falling back on patterns that have proven effective at other frontier-model shops. For the three research-heavy tracks (Models, Evals, Agents), candidates are assembling public evaluation write-ups: replication logs for published papers, ablation studies on open-weight models, benchmark dashboards they can walk an interviewer through line by line. A single well-documented notebook, showing data curation choices, training-curve diagnostics, and honest failure analysis, carries more weight than a list of conference proceedings, engineers who have interviewed at comparable labs report. The Evals role, in particular, rewards candidates who can demonstrate they have built custom harnesses rather than only running off-the-shelf suites.

On the Agent Infrastructure and Product tracks, the emphasis shifts to system-level artifacts. Candidates are shipping small but complete agent runtimes: a planner with tool-use loops, a memory layer with retrieval benchmarks, a deployment scaffold that handles versioning and rollback. Recruiters at peer firms say they now ask for a GitHub link before a résumé; a working prototype that survives a load test is treated as a de facto screen pass. The Founding Account Executive role, with its unusually wide band, appears to be filtering for operators who can show pipeline metrics from earlier zero-to-one go-to-market motions, not just quota attainment but evidence they defined the ideal customer profile, built the outbound motion, and hired the first two reps.

None of this is Jarmin-specific. The company's silence means candidates are preparing for a generic "frontier-tech bar" rather than a known Jarmin bar. That creates waste: some applicants over-index on research artifacts for an infrastructure role, others ship a polished agent demo when the screen is a distributed-systems design review. Until Jarmin publishes even a one-page "how we interview" note, the tactical arms race will stay decoupled from the actual target. The only verifiable data point remains the six role titles and their posted bands — everything else is pattern-matching against the broader market.

Competitors Copy the Homework

The frontier-tech talent market moves in packs. When a well-capitalized company signals a new hiring standard — especially one that de-emphasizes pedigree in favor of demonstrated output, peers tend to follow, not out of altruism but because the candidate pool itself recalibrates. Marc Andreessen described the dynamic bluntly in 2013: "Generally, number one is going to get like 90 percent of the profits. Number two is going to get like 10 percent of the profits, and numbers three through 10 are going to get nothing." The same concentration logic applies to talent. A firm that defines the screening bar effectively sets it for the tier below.

Jarmin's six new roles cover the full stack of applied AI deployment. The salary bands force competitors to justify their own compensation structures. But the deeper shift is in what Jarmin is screening for. A YouTube analysis of hiring-manager behavior, published in late 2025, captured the industry-wide pivot: "Hiring managers aren't hiring for your personality despite what TikTok says. They are buying outcomes." The same source laid out an eight-rule framework that has quietly become the de facto rubric at multiple frontier firms: stop memorizing trivia, study the business problem, describe delivery using the Google XYZ formula ("accomplished X as measured by Y by doing Z"), structure every answer, look intentional, be early, ask questions that reveal performance expectations, manage energy, and end like a consultant.

Competitors cannot ignore that framework when their own candidates start rehearsing it. Several firms in the applied-AI and robotics clusters have begun rewriting their interview rubrics to mirror the XYZ structure, replacing "tell me about a time you led a project" with "walk me through a metric you moved, the baseline, and the specific lever you pulled." The shift is visible in public job postings: requirements for "5+ years experience" are dropping in favor of "shipped a model to production serving 10,000-plus requests a day" or "built an eval suite that caught a regression before release." The language change is not cosmetic — it filters for the same hands-on problem-solving Jarmin's screen now prioritizes.

The consumerization-of-enterprise trend Andreessen identified a decade ago ("today all the consumerized enterprise stuff is as easily usable by the small business as it is by the large business") has a talent analog. Screening tools, interview platforms, and take-home environments that once lived only at hyperscalers are now accessible to Series A teams. That democratization means a 50-person company can run a screening loop that looks like Jarmin's: a live coding session against a real eval harness, a product-review exercise with actual user telemetry, a written memo defending an architecture choice. Competitors adopt these not because they've validated them independently but because candidates now expect them. A candidate who clears Jarmin's screen arrives at the next interview already fluent in the format; a firm that still uses whiteboard trivia looks obsolete.

There is also a defensive dimension. In winner-take-all markets, the cost of a false negative — rejecting a builder who would have shipped, exceeds the cost of a false positive. The Facebook API episode from 2013 offers a cautionary parallel: when the platform shifted to tiered thresholds to hide volatility, "sometimes the transparency made it tough for apps to make changes, because if their experiment failed and user counts dropped, the world would immediately know. This may have discouraged innovation." Hiring loops that over-index on clean narratives and brand-name résumés create the same chilling effect. Firms that tighten their filters to match Jarmin's outcome orientation are partly trying to avoid missing the next builder who lacks a FAANG badge but has a GitHub repo that proves the point.

The research does not name specific competitors reacting to Jarmin: no public statements, no leaked memos. But the structural forces are clear: a concentrated talent market, a visible shift in screening philosophy at a well-funded player, and a candidate population that moves between interviews carrying the new rubric like a passport. The firms that adapt fastest will be those that treat screening as a product problem — instrumented, iterated, and measured by the same XYZ logic they now ask candidates to demonstrate.

What One Hiring Burst Tells Us

Jarmin's addition of six roles in a single week offers a narrow but concrete window into how one frontier-technology company is allocating capital and attention. The first-party board data shows salary bands clustering for five technical roles and a wider range for the commercial role, with a board-wide median across salaried positions. Those numbers, current as of the past seven days, sit above the typical band the board tracks for comparable frontier-tech roles. Whether that premium reflects Jarmin's specific funding stage, the scarcity of the skill sets involved, or a deliberate signal to the market is not documented in the available research.

Those five technical titles map cleanly to the current architecture of large-language-model application development. Models and evals represent the upstream research and measurement layer; agents and agent infrastructure represent the downstream orchestration and tooling layer; product represents the integration layer. A company hiring across all five simultaneously is either building a full-stack AI product from scratch or expanding an existing one rapidly. The research does not disclose Jarmin's product, customer base, or revenue, so the strategic intent remains opaque. What is visible is the demand signature: simultaneous need for model-level expertise, evaluation rigor, agent-framework fluency, infrastructure engineering, and product judgment.

That commercial role's broad compensation range suggests Jarmin is also investing in early revenue motion, either landing design partners or converting pilots to contracts. In frontier tech, that hire often precedes or coincides with a shift from research-mode to commercial-mode. The board data does not indicate whether Jarmin has announced funding, launched a product, or signed reference customers. Without those signals, the broader market read is limited to: one company, six roles, compensation at or above board median, covering the full technical stack of LLM application development.

Extrapolating from a single company's hiring burst to sector-wide trends is hazardous. The research contains no competitor hiring data, no venture-capital deal flow figures, no talent-migration patterns, and no salary-survey aggregates beyond the board's own median band. Other frontier-tech firms may be expanding, contracting, or holding steady; the data does not say. What the board data does show is that Jarmin's six roles were posted within a seven-day window and carry compensation that clears the board's typical range. That is a fact. Whether it is a leading indicator, a lagging one, or an idiosyncratic data point cannot be determined from the available evidence.

For candidates and recruiters watching the board, the actionable signal is specific: roles titled "Member of Technical Staff" with domain modifiers (Models, Evals, Agents, Agent Infrastructure, Product) are being priced at base in San Francisco, and a founding commercial role carries significant variable upside. The technical domains map to skills that are currently scarce — evaluation methodology, agent orchestration frameworks, model fine-tuning or distillation, and the infrastructure to serve and observe agents at scale. Whether that scarcity persists, spreads, or collapses depends on forces — model capability jumps, open-source tooling maturity, enterprise adoption curves, that are not captured in this hiring snapshot. The screen Jarmin builds to filter for those skills will become the standard the next tier measures itself against, whether Jarmin publishes it or not.

Working in frontier tech? Zero G Talent tracks the openings: see every open Jarmin role, browse frontier tech jobs, the companies hiring, and the people building the field.