AI Researcher – RL & On-Device Models

Done training models that disappear behind an API? Put yours on phones, laptops and glasses instead. A San Francisco AI lab is building trustworthy, consumer-grade agents that run privately on the device. With their small action models and execution layer, an agent can read what's on screen and operate any app the way a person would, without APIs or custom integrations per app. They've held the top spot on the industry's leading mobile-agent benchmark since late 2025, work with a leading mobile chipmaker and a global device maker on edge deployments, and are backed by top-tier VCs and senior leaders from frontier labs. The founder has built and funded an agent startup before. The team has more than doubled in the past half year, and a new round is expected to close before the end of the year. Two lanes: consumer apps and partnerships with device makers. The core challenge: squeezing frontier-level capability into the tight compute and memory limits of a phone or laptop. What you'll ownLeading the training of a model family that powers the agents: pretraining recipe decisions, post-training (SFT, RLHF, DPO, GRPO and beyond), distillation, quantization, and every trick that helps a small model outperform its size One or more capabilities end to end: data mix → objective → evals → shipped into a production on-device runtime Your own workstream, measured on one clear metric, ending in a checkpoint that goes to production Experiments and write-ups the whole team builds on, and the discipline to drop ideas that don't move the needle Close work with infra and product engineers, the partnerships team and fellow researchers Your first 90 daysDay 30:you've reproduced a recent training run end to end and named the three bets with the most leverage Day 60:you're leading a workstream and have shipped a checkpoint that outperforms the best one so far Day 90:what you built is running in a partner's build What you bringHands-on depth across the training stack: pretraining, SFT/RLHF/DPO/GRPO, distillation and quantization Tricks for small models: distilling from frontier teachers, MoE at small scale, KV-cache compression, speculative decoding Experience with RL, on-device deployment, or both; systems fluency (memory, GPU optimization) is a big plus You design experiments around metrics that matter, and you ship checkpoints, not just papers You've lived through hypergrowth at a frontier lab, or you were a founder or one of the first hires at a startup that took off Relevance over pedigree: exceptional work counts more than a big-name logo