rlhf services hire AI safety researcher in Latin America

What does an AI safety researcher own on a US team?

Hire against a research question your team needs answered. A useful starting point is: which model behavior could cause harm, and what evidence would change a release decision?

For an RLHF project, ask the researcher to examine how human feedback affects the behavior you care about. Scope the work around a defined model, user group, and failure. A request to "make the model safer" leaves candidates guessing about the job.

Write the expected output into the role description. That might be an evaluation suite with documented limits, an experiment on preference data, or a report on failures under adversarial prompts. Name the person who will review it.

Keep research ownership distinct from release authority. The researcher should explain what the evidence supports and raise concerns. Your team should identify who accepts the remaining risk and who can pause a launch.

Define the work before choosing seniority

A junior hire should receive a bounded question and close review. Ask them to reproduce a result, inspect labels, or extend an existing evaluation. Give them access to a researcher who can resolve unclear cases.

A mid-level hire should be able to turn a scoped problem into an experiment. Test whether they can choose a baseline, detect weak evidence, and document a result another person can reproduce.

For a senior hire, test how they choose the question itself. Ask what they would stop measuring, which assumptions need challenge, and what finding would change their plan. Include responsibility for reviewing other people's work if the role requires it.

Do not make one role quietly absorb every task around RLHF. If the job includes managing raters, maintaining training infrastructure, or coordinating annotation delivery, say so. Candidates need to know how much time remains for research.

Give the researcher a usable working environment

Before interviews, write down what the hire can access. Include model versions, evaluation data, and approved compute. Assign an owner for access requests and sensitive findings.

For example, a researcher asked to assess harmful outputs needs an agreed way to handle those outputs. Set review and escalation rules before sharing the material. Explain the exposure involved during hiring.

Ask candidates how they would proceed if production data were restricted. Their answer should identify a useful first experiment and its limits. Avoid rewarding a plan that assumes unrestricted access to customer data.

Define what a completed assignment looks like before the hire starts. For an evaluation project, request the test cases, runnable code, and a memo explaining the result. Have another team member rerun the work before accepting it.

Include negative results in that agreement. A careful experiment may show that a proposed training change does not help. Judge the researcher on the method and the usefulness of the evidence, rather than requiring a favorable finding.

How does US pay from BLS OEWS compare with LatAm pay?

Use the US benchmark to frame the budget, then price the actual role. BLS reports a median annual wage of $135,980 for software developers in May 2025. Its Occupational Outlook Handbook uses OEWS wage data. This is a software developer benchmark, not a quoted salary for an AI safety researcher. BLS software developer pay

Do not apply a regional discount to that number and call the result a research salary. A role that owns research direction needs a different assessment from one that runs established evaluations.

For a Latin American hire, request a written quote against the same job scope and level used for the US search. Specify whether you want employee base pay, total compensation, or the full service cost.

Request a country rate card for the actual role

Use this rate card to collect monthly quotes. Keep the hours, currency, and employment model consistent across responses.

Salary context: Levels.fyi Mexico, Levels.fyi Brazil, and Stack Overflow Developer Survey 2025. These are distinct measures; do not treat them as a role-specific service quote.

Published LatAm AI safety researcher ranges corroborated by two public salary benchmarks: Not published.

Levels.fyi describes total compensation as base salary plus the yearly value of stock grants and bonus. Keep that definition attached to any comparison. A base-pay offer and an equity-inclusive figure answer different budget questions. Levels.fyi compensation

Ask for the currency in which the employee is paid and the currency on your invoice. Have the agreement explain how exchange-rate changes affect the bill. Also ask how bonuses, leave, equipment, and compute are treated.

The Opus salary guide distinguishes local salary from a full service price. This site is an Opus-owned reference; recommendations are editorial.

Opus reports that it charges one flat monthly rate per hire. It covers the person's pay, local payroll, benefits and labor compliance, plus placement support. There is no separate recruiting fee. The rate depends on role, seniority and location. Published dollar rates: Not published. Opus pricing

Request the role-specific quote before comparing totals. Keep research compute and any separate data purchases visible in your own budget.

For an hourly proposal, ask how the supplier bills meetings, failed experiments, and time spent waiting for access. For a monthly proposal, confirm the expected working hours and any exclusions. Record these terms beside the price so a lower headline rate does not hide a smaller scope.

Set a budget owner for experiments too. Give the researcher a clear spending limit and a way to request more compute. That lets your team review a proposed run before it creates an unexpected bill.

Which skills should you test, and how?

Use a paid work sample that resembles an early assignment. Keep it small enough to review closely. Give each candidate the same task, access, and written expectations.

One useful exercise is to provide a small set of model responses with a draft safety rubric. Ask the candidate to critique the rubric, propose an evaluation, and produce a short analysis. Use synthetic or approved data.

The exercise should leave room to disagree with the task. A candidate who finds a flaw in the metric should be able to explain it without guessing what the interviewer wants.

Ask the candidate to walk through the work live. Change one assumption, such as the user population or labeling rule. Watch whether they update their conclusion or defend the original answer without checking.

Test the research behind the result

A score increase alone should not settle the interview. Ask whether the evaluation data influenced the method. Have the candidate explain how they would detect leakage, duplicates, or changes in the test set.

For preference work, ask what should happen when raters disagree. Look for a reasoned distinction between unclear instructions and valid differences in judgment. The candidate should explain when to preserve disagreement rather than collapse it into a single label.

Ask how they would compare a safety improvement with a loss in useful answers. Give them a concrete example where refusing every request would score well on a narrow test. Have them redesign the test.

For code, request enough detail that another team member can rerun the result. Review how the candidate records model versions and handles failed requests. A polished chart should not conceal an unstable experiment.

Include a case where the average result hides a failure for a smaller user group. Ask the candidate which examples they would inspect and how they would report the uncertainty. Look for a justified next step rather than a confident claim based on too few cases.

Agree on permitted tools before the exercise. If candidates may use AI coding tools, ask them to explain and check the resulting code. Assess whether they understand the analysis well enough to catch an error during review.

Assess seniority through decisions

For a senior candidate, discuss a proposed research agenda. Ask which experiment they would run first with limited compute and which result would make them stop. Probe the tradeoff they chose.

For a junior candidate, focus on whether they can follow a method, spot an inconsistency, and ask for review. Do not use an open-ended leadership exercise to reject someone hired for supervised work.

Agree on the scoring rubric before interviews. Have reviewers write their assessments independently, then discuss differences. Record the evidence behind the decision so the hiring manager can explain it.

If the work uses an existing RLHF platform, include a task that fits its interfaces. Assess the candidate's reasoning separately from their familiarity with that particular tool.

Talk to Opus about an AI safety researcher role.

Where should you search for Latin American talent?

Start with the working schedule and research requirements. Include Mexico, Colombia, Argentina, Brazil, and Costa Rica in the search where those constraints fit. Ask for candidates whose work samples match the role.

Choose people through evidence of their research. Country-level comparisons should not replace a review of code, methods, and communication. Ask each candidate to explain their own contribution to any shared project.

For a publication, request a discussion of the method and its weaknesses. For private work, accept an anonymized explanation or a fresh exercise. Do not make candidates disclose a former employer's confidential data.

Compare overlap using the candidate's city

The table shows shared hours when both teams work 9 a.m. to 5 p.m. locally. Each pair lists US standard time first, then US daylight time. It is a scheduling calculation, not a promise about a candidate's availability.

Time zones and overlap calculations: IANA North America data and IANA South America data. The seasonal examples use January 15, 2027, and July 15, 2026.

Mexico City and San José stay at UTC-6. Bogotá stays at UTC-5. Buenos Aires and São Paulo stay at UTC-3. Those cities do not change clocks for daylight saving time. Tijuana follows US daylight saving changes, so confirm the city rather than using a single schedule for Mexico. IANA North America, IANA South America

An Eastern team and a Pacific team may reach different decisions about the same candidate. Put the required review window in the job description. Discuss any schedule shift with the candidate before the offer.

Keep live meetings focused on decisions and unresolved findings. Use written experiment notes for routine updates. Ask how the candidate would flag a concern when the US manager is offline.

Check communication in the work itself

Use the language your team uses for technical decisions during the assessment. Ask for a written conclusion, then discuss a challenge to it. Assess whether the candidate can explain uncertainty without hiding behind jargon.

If the research covers Spanish or Portuguese outputs, test that work directly. Ask the candidate to explain ambiguous wording and when a native-language reviewer is needed. Residence alone should not establish language or domain expertise.

Before signing, confirm the work location, agreed schedule, and holiday coverage. Ask who will handle urgent findings during leave. Put that responsibility with a named person.

Which hiring route fits: recruiter, EOR, augmentation or managed team?

Choose the route based on who will direct and review the research. If your team cannot assess the work, adding a researcher will not resolve that management gap.

Use the same role description when comparing models. Otherwise, a quote for a person and a quote for a delivered evaluation may look comparable while covering different work.

For business process outsourcing services that include labeling, define the boundary between annotation delivery and research judgment. Ask who writes the rubric, resolves disputed labels, and checks whether the data supports the experiment.

When comparing RLHF companies, request the names and relevant work of the people who would perform the research. Ask who reviews their methods. A description of annotation capacity should not substitute for that assessment.

For a managed assignment, write acceptance criteria around the work you can inspect. Require access to experiment records and enough documentation to reproduce the findings. Agree on what happens if the result is inconclusive or the original question needs revision.

For an individual placement, plan continuity before access is granted. Keep code and experiment notes in company-controlled systems. Name someone who can take over a running study if the researcher leaves. Ask the hiring partner how a replacement search would work, then check those terms in the agreement.

Confirm how Opus would fit your team

Opus reports that the client chooses the finalist and manages day-to-day work. Use that model when your team can set priorities and review research results. Opus pricing

Opus reports that every candidate passes an 8-layer vetting process and every placement completes a 6-module internal AI certification before day one. Ask how that screening maps to your research exercise. Keep your own technical review in the hiring process. Opus hiring process

Separate hiring speed from readiness to work

SHRM describes median time-to-fill in 2025 as roughly a month and a half for executive and nonexecutive positions. It measures the period from job requisition to offer acceptance. Treat that as broad recruiting context rather than a forecast for this specialist role. SHRM recruiting benchmark

Opus reports 3 vetted candidates in 5 days. It also reports about 14 days from brief received to hired, not first day worked. Reporting period and calculation: Not published. These milestones use a different starting point from SHRM's measure. Opus hiring process

Build your start plan around the accepted offer, notice obligations, and access setup. Name the first research question and its reviewer before day one. Have an approved dataset ready so the hire can begin useful work.

Send the hiring partner a role description with the research problem, expected output, reporting line, and required shared hours. Include the assessment process and budget basis. Request a written scope and monthly quote against those terms.

Talk to Opus about an AI safety researcher role.

<!-- File not saved: filesystem access is read-only. Automated page validation was not run. -->