VBT Task Author

Careerflow Human Data Labs partners with AI companies to bring real-world professional expertise into their products.

Role Overview

We are seeking an analytical researcher and red-teaming specialist to build evidence-based web-research benchmark tasks that reliably expose weaknesses in state-of-the-art (SOTA) AI models (e.g., Claude Opus 5 Max). In this role, you will discover hard-to-surface facts, build precise research tasks around them, test them directly against leading LLMs, and diagnose the exact failure mechanisms in the model's search and reasoning paths.

Your work will directly shape AI evaluation frameworks, ensuring next-generation frontier models are tested against rigorous, non-trivial real-world research scenarios.

Job Requirements:

• Core Experience: 1+ years in investigative research, fact-checking, AI benchmark task authoring, red-teaming, or technical QA review.

• Benchmark Familiarity: Proven experience authoring, testing, or reviewing tasks for AI benchmarks such as GAIA, BrowseComp, SimpleQA, SWE-bench, Terminal-Bench, or Humanity’s Last Exam (HLE).

• Analytical & OSINT Skills: Intermediate-to-advanced web search and Open Source Intelligence (OSINT) skills; ability to navigate academic archives, local records, or obscure databases.

• Model Trajectory Debugging: Strong analytical capability to inspect model reasoning chains, search logs, and source selection to diagnose root causes of failure.

• Precision Writing: Ability to author clear, unambiguous benchmark prompts with bulletproof, primary-sourced answer keys.

Job Responsibilities:

• Obscure Fact Discovery: Deeply research niche topics to identify factual ground truths that cannot be surfaced via simple, direct search queries.

• Source Verification: Rigorously verify facts against primary or highly credible sources, cross-check for contradictory evidence, and document exact source URLs.

• Task Construction: Draft clear, answerable benchmark tasks designed specifically to expose structural weaknesses in model research strategies.

• Red-Teaming & Model Testing: Execute tasks against designated SOTA models to confirm they reliably trigger intended model failures.

• Failure Diagnosis & Debugging: Trace model run logs, search trajectories, and intermediate outputs to pinpoint exactly where and why the model failed (distinguishing genuine model limitations from ambiguous prompts).

• Pattern Scaling & Review: Apply validated "model-breaking" research patterns across new domains and, for strong performers, review tasks authored by peer trainers.

Preferred Qualifications

• Bachelor or higher degree from relevant engineering domain is a must for this position.

• Prior experience specifically designing prompt-injection or reasoning-trap tasks that break SOTA models.

• Advanced OSINT capabilities or published investigative/academic research demonstrating deep source verification.

  Location

  Bangladesh, India, Indonesia, Egypt, Ghana, Nigeria, Turkey ,Brazil, Columbia

  Total number of positions: 180

  Engagement Length: 4 weeks, Full-time (8 hours/day) 40 hours per week. Overlap 4 hours with PST