Research Engineer, Post-Training

Actively AI•New York, NY

22h

About The Position

Actively AI is seeking research engineers to accelerate post-training models for the go-to-market and sales domain. Its platform deploys Per-Account Agents across customers’ total addressable markets to research, identify opportunities, and advance next steps around the clock. The product supports sales representatives and leaders across prospecting, deal execution, expansion, and retention, with demand for higher-quality intelligence across these tasks. The role works directly with the founders to fine-tune model intelligence and advance performance in the domain. The engineer will develop post-training recipes, design meaningful evaluations, investigate unexpected training results, apply alignment methods to shape agent behavior, and study how performance scales with data and compute.

Requirements

A track record of advancing machine-learning systems through post-training, alignment, or related methods such as RLHF, RLAIF, preference modeling, or reward learning.
Strong fundamentals in probability, statistics, and machine-learning theory.
Ability to distinguish genuine effects from noise and bugs when analyzing experimental data.
Ability to understand end-user tasks and domain context and develop intuition for judging what good, great, and poor performance look like.

Nice To Haves

Evidence of original contributions or experience working on similar problems in industry.

Responsibilities

Iterate on datasets, training stages, and hyperparameters that determine model behavior, and measure their combined effects on evaluations and production performance.
Design, optimize, and rebuild evaluations that capture meaningful model performance, and ensure evaluation results are valid and informative.
Investigate unexpected training results to understand their causes and carry those insights into future work.
Apply and advance RLHF, RLAIF, and constitutional approaches to shape how agents reason, act, and collaborate with humans in long-horizon tasks.
Measure how performance scales with data and compute, and develop new methodologies when existing approaches reach their limits.
Develop a strong understanding of the end-user tasks and domain, and use user empathy and intuition to judge the quality of model outputs.