Careers at Variance: Teams, Pay and How to Get Hired

Who gets hired and onto which teams

Business Wire reported that Variance raised $21.5 million in its March 2026 Series A announcement to build AI agents for risk investigations. The company described itself as a “small, talent-dense team” tackling problems that touch billions of events and entities. Its listed roles span San Francisco and New York, while its careers page says employees take end-to-end responsibility and that it measures prototype-to-production timelines in days, rather than months.

The operating model is deliberately hands-on. That points to candidates who can make decisions inside unfinished work, connect technical choices to a business result and proceed without needing each step assigned.

The company said on its careers page:

“Every employee at Variance is an owner from day one. Ownership here goes beyond titles, as we trust our employees with full, end-to-end responsibility for the systems we build and the outcomes we deliver to the world.”

Karine Mellata and Michael Lin founded Variance in 2023 after they met while working on Apple’s fraud engineering and algorithmic risk team. The company was formerly called Intrinsic. Variance focuses on artificial-intelligence agents for investigation and decision-making in regulated risk workflows.

Variance said large Fortune 500 companies and financial technology companies use its agents for KYC, KYB, AML, transaction monitoring and fraud detection. Those investigations typically require teams of analysts and days to complete; Variance says its agents can return decisions in minutes.

In company-reported results, Variance said its agents achieved human-level precision across hundreds of know-your-customer investigations at a leading Fortune 500 company and replaced manual review queues. At one of the world’s largest retailers, its agents traced chains of ultimate beneficial ownership, uncovered coordinated fraud rings and cross-referenced sanctions and adverse media in minutes rather than weeks. Because Variance produced those results, they remain company claims rather than independent benchmarks. The work calls for someone who can reason through uncertain evidence and connect model behavior to a compliance decision.

The technical core around judgment

The board’s four technical openings are all based in San Francisco: Research Engineer, Evals, Research Engineer, Judgment Systems, Senior Software Engineer and Forward Deployed AI Engineer.

The first two titles point to evaluation and judgment systems, while the latter two point to software engineering and forward deployment. The available board records do not reveal reporting lines, the tool stack, experience requirements or full responsibilities. They identify four functional emphases—evaluation, judgment systems, software engineering and forward deployment—but do not show how those functions are divided or sequenced.

Commercial roles and the two-city structure

Variance is also hiring for commercial work. In San Francisco, its listed openings include Enterprise Account Executive and Product Marketing Lead. The board also lists Business Development Representative and a New York-based Enterprise Account Executive. These roles place customer acquisition, product positioning, business development and enterprise sales alongside research and engineering.

The location tags narrow the search. Among the listed roles, the New York opening is for enterprise sales, while the technical openings are in San Francisco. The public records support a two-city hiring footprint but do not establish a New York engineering center. Public sources do not provide reporting-line data, so the openings reveal recruiting priorities, not an org chart. The mix spans research, forward deployment, product marketing and two commercial functions: enterprise sales plus business development.

What it pays

The first-party Variance job-board data reports a typical annual band of $70,000 to $192,000, with a $160,000 median. The board says the summary reflects two salaried roles, so it describes the available dataset rather than a companywide salary policy or a prediction of where an offer will land.

The table preserves what each posting says about pay and marks missing data instead of estimating it.

Only two of the listed postings publish a numeric range. The forward-deployed range starts above the business-development range, and the two overlap only narrowly; each also extends beyond one edge of the board’s summary band. The median therefore describes a limited dataset, not a typical Variance offer.

Most technical postings and both enterprise-sales listings state no numeric band. The available data cannot show whether function or location affects pay.

Beyond salary, Variance’s careers page lists unlimited paid time off, sick leave and parental leave, along with medical, dental and vision coverage for employees and their dependents. It also lists a 401(k) program and a monthly stipend that can cover gym, fitness, mental health or wellness expenses. The page gives no 401(k) match formula or paid-leave limits, so candidates must ask for those terms.

The published information is thinner on equity. Variance said the package includes “meaningful equity,” but the careers materials and board data do not state what the grant contains or how many shares it covers. The available sources provide no valuation and do not explain vesting, exercise terms or refresh policy. A candidate therefore cannot calculate the grant’s dollar value from the public materials.

Before accepting, ask the recruiter to put the role’s salary band, level, target bonus or commission structure, equity grant terms and 401(k) match in writing. In a small, talent-dense team, an incomplete offer leaves too much uncertainty to interpret from career-page copy alone.

How the hiring process works and what gets candidates through it

Variance has not published a stage-by-stage map of its hiring process. Its live Variance page lists openings, but the available records do not identify the number of interviews, who sits on the panel, how answers are scored or what triggers rejection. No public source supports a guessed sequence.

The records nevertheless show different functional emphases. The titles cover evaluation, judgment systems, software engineering, forward deployment, product marketing, enterprise sales and business development. The board excerpts do not disclose full duties, required experience or the screening method, so they cannot support a guessed interview loop.

For applicants, the most defensible way to show ownership is to describe what they personally decided, built and delivered. A résumé that lists technologies without showing that decision-making offers little evidence of research or engineering ownership.

What a recruiter screen can and cannot tell you

No available material confirms a recruiter stage. Candidates should ask whether one is used and, if so, which role the company is hiring for, how the team defines success, what the interviews assess and whether the process differs by location. Those details must come from the company rather than an assumed standard tech-company loop.

An application can be made legible without guessing at a hidden filter. Lead with the opening that fits, then name a shipped system or concrete result. A reviewer can then assess relevance from what you owned rather than from a list of tools.

Showing technical depth instead of naming tools

No public source says Variance uses a coding test, machine-learning exercise or take-home project. That absence makes the substance of a candidate’s examples more important, not less.

DataThink’s data-science interview guide says the definition of a model problem only starts the test; a candidate must inspect behavior, identify the failure and choose a fix that fits it.

That standard maps well to the work signaled by the evaluation and judgment openings. “Use XGBoost” does not show whether a model underfits or overfits. A stronger answer notes that training and validation errors are both high, then explains the model change or feature work that would test the diagnosis. For evals and judgment systems, the same principle applies outside model training: describe the failing behavior, connect the remedy to it and state what evidence would show whether the fix worked.

FloCareer, after more than one million interviews across over 600 organizations, found that evaluations can vary with the structure of the interview, who conducts it, expectations, scoring calibration and time pressure.

FloCareer said:

“Standardized competencies, clear scoring criteria, and consistent formats reduce arbitrary variation and make evaluations more comparable.”

That research does not describe Variance’s process. It does suggest a sound tactic: ask which competencies the interviews assess and what work counts as evidence.

Strong signals and disqualifiers

Variance publishes no formal disqualifiers, and no source identifies an automatic rejection. Its careers page does state that employees take responsibility from end to end. For applicants, describing personal decisions and outcomes is therefore more informative than listing technologies. DataThink’s diagnostic standard adds another test: a recommendation should connect to the observed failure rather than stop at a tool name.

The most useful application evidence to prepare is an artifact whose full chain is visible: the problem, the constraints, the decision, the result and what the candidate would improve with more time. For research work, make the failure mode and evaluation method explicit; for software and deployed roles, make the shipped system and its operating consequence clear. Choose the artifact most relevant to each opening, then ask for the scoring criteria and interview format rather than inventing steps the public record does not support.

Where the work happens

The live Variance page locates the company by city, not by room. Public sources identify office addresses and in-person roles, but they do not show how the space is used or whether customer work happens on site.

aVenture’s company profile, updated October 19, 2026, lists Variance at 570 Pacific Avenue, Floor 3, San Francisco, California 94133. The board’s San Francisco postings place research and engineering in the city, but they do not reveal how many people share the floor or how the space divides.

Variance’s LinkedIn company profile lists another location at 564 Market Street, Suite 623, San Francisco, California 94105. The aVenture record is the newer dated record, but neither source explains the relationship between the addresses, whether the Market Street suite remains active or which functions occupy either location. No board posting assigns a team to a specific suite, so candidates should confirm the work site before planning a commute around either address.

Public headcount data conflicts. The more recently dated aVenture profile lists four employees; LinkedIn gives a range of 11–50. Neither source divides those figures by location, so neither can support a precise claim about floor density or reporting structure.

They also say little about where customer integration happens.

A recruiting post for Agent Product Manager and Agent Engineer listed both roles as in-person in Jackson Square and said:

“You'll be in the room with risk teams at the largest banks, fintechs and online platforms in the world, watching them work high-stakes workflows and scaling verticalized agents.”

That places customer-facing work at the center of these roles. Variance’s company profile says its agents work from alert to decision, and its product page describes evidence collection and complete audit trails. The recruiting post frames engineers and product managers as working directly with risk teams; it does not identify a dedicated company facility for that work.

An office footprint, not a public lab network

No available Variance page or board posting describes a robotics lab, machine shop, prototype workshop, secure investigation center or hardware test bench. The published material presents a software company, not a laboratory operation.

For a candidate for the Forward Deployed AI Engineer, deployment questions are material. The available public summary does not include the full job description. Variance says its systems are used in production at Fortune 500 scale, and its recruiting post refers to major banks, fintechs and online platforms. Candidates should ask whether the role involves customer sites and how work moves between those sites and Variance offices. Public materials say Variance supports fully on-premise and air-gapped deployments with customer-managed encryption keys, field-level data controls and confidential computing, but they do not describe this role’s travel or customer-review arrangements.

It remains unclear whether employees travel to client offices, join customer reviews remotely, rotate between sites or remain at Variance offices. The footprint therefore says more about where the company hires than how its customer-facing work is organized.

Who thrives here

Variance’s careers page names four operating values: Impact, Ownership, Velocity and Craftsmanship. A similarly named Team Variance is a different organization; its principles should not be attributed to Variance.

Impact calls for full engagement; Ownership gives every employee end-to-end responsibility; Velocity emphasizes speed from prototype to production in days, not months; Craftsmanship asks for care and intentionality in mission-critical systems. Read together, they favor people who explain choices, surface risk early, accept correction and keep the next step moving. Ownership covers the systems and outcomes, not just a title.

The product makes that standard concrete. Variance’s agents reason over complex, fragmented data and return auditable decisions under each customer’s compliance procedures, so people working on these systems must understand case context and notice where evidence is weak. They must also treat a wrong decision as a business and human cost. A candidate example should frame the problem, test system behavior, explain the consequence and own the result rather than only generate an answer.

That creates an important distinction: independent does not mean isolated. Variance said it wants every team member “fully-engaged in fulfilling our mission.” A candidate can show this by choosing a method, explaining trade-offs, asking for missing domain information and proceeding when perfection is unavailable.

Candor is a condition, not a style trick

MIT Sloan Management Review’s Great Resignation analysis examined 1.4 million Glassdoor reviews and found that toxic culture was the strongest predictor of industry-adjusted attrition, 10.4 times as predictive as compensation. Workers who felt disrespected and saw unethical behavior were among the leading elements of toxic cultures. For Variance, whose agents produce auditable decisions, candor about uncertainty and error matters; candor without respect is another failure mode.

Only four Glassdoor reviews appear in the available record. It provides the count but not the review text, so it cannot settle questions about daily office life or serve as a personality test.

Velocity without craftsmanship is not enough. Variance lists both, along with Ownership and Impact. Harvard Business School’s employee-engagement guidance links poor workplace communication with delayed projects, lower morale and missed performance goals.

Applicants can show ownership by making the state of the work legible: naming what is known, what remains uncertain, where help is needed and what happens next. They can disagree with a decision, document the reason, commit to a path and own the result. That preserves speed without sacrificing context.

Autonomy needs context

Freedom should not be confused with freedom from accountability. A study of 1,832 respondents in China found that the effect of working-time variation on self-reported health depended on autonomy. The study associated variable hours with better health when autonomy was high and worse health when autonomy was low.

Variance’s emphasis on ownership points toward a similar condition: autonomy needs context, not freedom from accountability. Harvard Business School’s guidance pairs delegation with transparent communication, respect and learning. Without those supports, autonomy can weaken accountability and increase risk.

A candidate can test that condition in an example: did they surface risk before it became a failure, explain what remained uncertain and revise the work after a disappointing result? Those details show whether independence included accountability, not merely solo work.

The honest fit test

Even with the operating values, fit is not something a values page can settle. The operating model may suit you if you get energy from carrying a problem across functions and can defend a decision without making it personal. It may strain people who need a large support organization or measure progress by title and process rather than completed work.

Before applying, test the fit against your own record. Ask whether you can spend enough time in a customer’s context to make a sound call, remain with the problem after the obvious fix and pass on information the next person can use.

Among the board’s listed roles, San Francisco carries the technical openings and New York carries an enterprise-sales role. The work behind those openings is not small: an uncertain investigation must become a defensible decision. The candidate who can show how they made that turn—without waiting for every step to be assigned—has supplied the evidence this guide recommends surfacing.

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