Part-Time Government & Public Policy Subject Matter Expert (Remote)

Company Description

Halluminate is an applied AI and data company working with frontier labs to train financial services and knowledge work superintelligence. We build high-quality computer-use and tool-use reinforcement learning (RL) environments that train AI agents to do real work across financial services verticals and adjacent domains.

Role Description

Halluminate is hiring part-time Government and Public Policy Subject Matter Experts (SMEs).

Frontier AI models can already build a clean budget model, run a standard variance analysis, and reconcile a simple ledger. We need experts who can push them further. In this role you'll design work that would challenge a strong senior analyst or controller, with multi-step reasoning, messy inputs, and the correct judgment calls where the right answer isn't obvious, then use it to expose where today's best AI models still fail. Your work feeds directly into the training data and frontier evaluations that leading AI labs use to measure and improve their models.

Your day-to-day in this role is RL environment task development and writing (i.e., creating the complex problem and solution that the AI agents must attempt to solve), not data labeling AI outputs based on your preferences or something less challenging.

Responsibilities

• Shape the future of knowledge work by identifying which financial workflows matter most and where today's AI models fall short, drawing on your real experience
• Build training materials for AI systems, including policy briefs, government program cost analyses, budget justifications and regulatory impact assessments, and other real-world deliverables
• Create challenging tasks that expose weaknesses in state-of-the-art AI models on financial services work
• Contribute to frontier evaluations that measure how leading AI models perform on real finance workflows
• Work directly with research and engineering teams at frontier labs to evaluate model outputs and turn findings into better training data
• Revise tasks quickly based on feedback from reviewers and model results, often across multiple rounds

What we're looking for

• 2-5+ years of experience working in government or public policy (federal, state or local government, public sector consulting, policy analysis, etc.)
• Exceptional command of core workflows in Excel, PowerPoint, and other professional software programs (modeling, valuation, deal materials, memos)
• Curiosity about machine learning, AI, data, and engineering concepts. No technical background required.
• Comfort with direct, frequent feedback and the ability to turn it around quickly without losing quality
• Sharp attention to detail and the judgment to spot exactly where an output would fail a senior review
• Strong organization and full ownership of deadlines

Location

• Fully Remote

Time Commitment

• Minimum of 20 hours per week
• You choose when to work

Compensation

• Training milestone payments between $500-$4000
• Implied hourly rate: $100/hour - $200/hour
• Bonuses paid out for above average performance or stellar contributors

If interested please apply here:

Why work with Halluminate

Halluminate builds the highest quality, most complex training environments and benchmarks that leading AI labs rely on to push their most capable models forward. Our expert researchers lead that work. They design the problems, set the standards models are measured against, and uncover where frontier AI still falls short. Where other competitors may focus on data labeling / annotation, our experts are challenged to identify weak points in frontier labs models to accelerate AI capabilities in finance.

• Market leading compensation. Our experts earn among the highest rates in the industry, with median earnings of $3,800 in September and our top 25% averaging over $30,000 that month.
• Consistent Demand for Work. Halluminate experts get a steady stream of new projects, not stop-start bursts. You choose how much to take on, based on your schedule and earning goals.
• Fully remote, you set your schedule. All work on the platform is completed remotely. We support work from outside the US: if you're traveling abroad temporarily, flag it to your Team Lead (note that work is not permitted from certain countries), and if you'll be abroad for an extended period, we can set your start date for when you return. Experts can work when they want and where they want as long as they hit their 20 hours per week minimum. There are no meetings, Zoom calls, or Slack huddles. All feedback from reviewers is left on our internal platform
• Build the environments frontier models learn from. You design RL tasks grounded in real data rooms, define the expert-quality answers models are rewarded against, and find the failure modes that show where today's models fall short of professional judgment. Your work feeds directly into how leading AI labs train and evaluate their models.
• Push the frontier of what AI can do. Our experts have contributed to industry benchmarks like Westworld Finance Diligence Bench, which tests AI agents on a complete acquisition due-diligence process.
• A path into AI research. Top contributors have moved into team lead and full-time research roles at Halluminate.
• Join an elite network. We work with a select group of professionals that are capable of building these AI training environments. You will be among a prestigious cohort with dedicated support in advancing the future of knowledge work.

Our hiring process

Four steps, each a preview of the actual work. You'll know quickly if it's a fit, and so will we.

1. Resume Review. We're looking for real depth in most knowledge work domains so please apply. We may have a project that fits your background now or in the near future.
2. Proctored Exam (60 minutes). Show us how you think through problems from your field.
3. Screening Call (30 minutes). Dive into your experience working in your field and with AI.
4. Mock Problem Build (3 business days). Build a task like the ones you'd design with us. We provide the inputs and do not use any of the outputs, this is purely evaluative.