AI Product Operator, Niural AI Labs ( Night Shift )

About Niural At Niural, we're not just building products; we're redefining how companies operate in a new era of the internet driven by AI. We're creating truly global solutions that empower businesses to thrive in the digital economy. If you're eager to join a relentlessly ambitious team, develop tools that will influence the flow of billions of dollars, and share in the asymmetric financial rewards of building foundational internet infrastructure, then come help us shape the future of finance. We are a global Payroll, Employer of Record (EOR), Agent of Record (AOR), and Contractor Management platform.

Niural AI Labs sits inside our product team, not off to one side of it. We are building the Learning Gym : the environments, reward functions, and evaluation harness where our AI agents are trained and measured on real Niural work. Payroll runs, contractor onboarding, invoice reconciliation, compliance checks. Work with actual right answers. That last part is why this is a good research problem. Most agent environments have to fall back on a model judging another model's output, which drifts and can be gamed. Ours does not. Correct gross to net pay is checkable. A missed filing deadline is checkable. You can build reward signals from ground truth rather than from opinion, and that is rare enough to be worth writing about. Which is the other half of the job. You will write research papers for Niural AI , first author, out of the work you do here. We are not looking to republish anything you have already done. Your existing publication record is simply how we know you can take a result from idea to defensible write-up, because most engineers cannot. You need to be a generalist. In one month you might build an environment, wire up the training loop, write the dashboard the team watches it on, and draft the paper. Nobody will hand you a spec, and nobody else will do the parts you find boring.

Build the Learning Gym ● Build sandboxed environments that wrap real Niural workflows, stateful across episodes so that an action in one task changes the state a later task sees ● Seeded and replayable, so a result from six weeks ago can be reproduced today ● Difficulty knobs and curriculum structure, so environments can stay at the edge of what our agents can currently do ● Realistic mess as a feature, not an accident: missing fields, stale records, contradictory sources. Payroll data supplies all of it for free ● Generate the synthetic data these environments run on, because real customer payroll data cannot leave our systems Engineer the rewards and verifiers ● Design programmatic verifiers from known ground truth, and reach for a model as judge only where nothing else is possible ● Score trajectories rather than final answers. Two runs can take different paths and both be right, or be wrong for different reasons ● Build against reward hacking and against contamination, and assume the agent will find whatever hole you leave Train agents, and prove it worked ● Run the training: reinforcement learning loops, curriculum schedules, fine tuning and distillation where they earn their place ● Establish honest baselines, run ablations, and measure whether an improvement generalizes or only fits your environment ● Publish negative results internally. A well measured failure is worth more to us than a demo Prototype fast ● One to two week spikes on agent architectures, memory strategies, tool use patterns and scaffolds, tested against the Gym rather than against your intuition ● Hand what works to the teams building our production agents, and stay involved until it lands Write ● First author research papers for Niural AI, taken from real work, with reproducible methodology and an honest limitations section ● Internal technical reports that the product and engineering teams actually read ● Represent the work externally over time: preprints, talks, and possibly open sourced environments Who