Together AI is an AI neocloud that provides AI-focused cloud infrastructure and tools to support adoption of open-source models.
<h3><strong>About the Role</strong></h3> <p>This is a research engineering role with direct production impact. You won’t be publishing ideas in isolation—you will translate new RL algorithms, scheduling methods, and inference optimizations into production-grade systems that power Together’s API. Success in this role means shipping measurable improvements in latency, throughput, cost, and model quality at scale. We are looking for researchers who enjoy owning systems end-to-end and turning frontier ideas into robust infrastructure.</p> <p>The Core ML (Turbo) at Together AI team sits at the intersection of efficient inference (algorithms, architectures, engines) and post‑training / RL systems. We build and operate the systems behind Together’s API, including high‑performance inference and RL/post‑training engines that can run at production scale.</p> <p>Our mandate is to push the frontier of efficient inference and RL‑driven training: making models dramatically faster and cheaper to run, while improving their capabilities through RL‑based post‑training (e.g., GRPO‑style objectives). This work lives at the interface of algorithms and systems: asynchronous RL, rollout collection, scheduling, and batching all interact with engine design, creating many knobs to tune across the RL algorithm, training loop, and inference stack. Much of the job is modifying production inference systems—for example, SGLang‑ or vLLM‑style serving stacks and speculative decoding systems such as ATLAS—grounded in a strong understanding of post‑training and inference theory, rather than purely theoretical algorithm design.</p> <p>You’ll work across the stack—from RL algorithms and training engines to kernels and serving systems—to build and improve frontier models via RL pipelines. People on this team are often spiky: some are more RL‑first, some are more systems‑first. Depth in one of these areas plus appetite to collaborate across (and grow toward more full‑stack ownership over time) is ideal.</p> <h3><strong>Responsibilities</strong></h3> <ul> <li><strong>Advance inference efficiency end‑to‑end</strong></li> <ul> <li>Design and prototype algorithms, architectures, and scheduling strategies for low‑latency, high‑throughput inference.</li> <li>Implement and maintain changes in high‑performance inference engines (e.g., SGLang‑ or vLLM‑style systems and Together’s inference stack), including kernel backends, speculative decoding (e.g., ATLAS), quantization, etc.</li> <li>Profile and optimize performance across GPU, networking, and memory layers to improve latency, throughput, and cost.</li> </ul> <li><strong>Unify inference with RL / post‑training</strong></li> <ul> <li>Design and operate RL and post‑training pipelines (e.g., RLHF, RLAIF, GRPO, DPO‑style methods, reward modeling) where 90+% of the cost is inference, jointly optimizing algorithms and systems.</li> <li>Make RL and post‑training workloads more efficient with inference‑aware training loops—for example, async RL rollouts, speculative decoding, and other techniques that make large‑scale rollout collection and evaluation cheaper.</li> <li>Use these pipelines to train, evaluate, and iterate on frontier models on top of our inference stack.</li> <li>Co‑design algorithms and infrastructure so that objectives, rollout collection, and evaluation are tightly coupled to efficient inference, and quickly identify bottlenecks across the training engine, inference engine, data pipeline, and user‑facing layers.</li> <li>Run ablations and scale‑up experiments to understand trade‑offs between model quality, latency, throughput, and cost, and feed these insights back into model, RL, and system design.</li> </ul> <li><strong>Own critical systems at production scale</strong></li> <ul> <li>Profile, debug, and optimize inference and post-training services under real production workloads, taking research ideas all the way to stable, measurable improvements in deployed systems.</li> <li>Drive roadmap items that require real engine modification—changing kernels, memory layouts, scheduling logic, and APIs as needed.</li> <li>Establish metrics, benchmarks, and experimentation frameworks to validate improvements rigorously.</li> </ul> <li><strong>Provide technical leadership (Staff level)</strong></li> <ul> <li>Set technical direction for cross‑team efforts at the intersection of inference, RL, and post‑training.</li> <li>Mentor other engineers and researchers on full‑stack ML systems work and performance engineering.</li> </ul> </ul> <h3><strong>Requirements</strong></h3> <p>We don’t expect anyone to check every box below. People on this team typically have deep expertise in one or more areas and enough breadth (or interest) to work effectively across the stack. The closer you are to full‑stack (inference + post‑training/RL + systems), the stronger the fit—but being spiky in one area and eager to grow is absolutely okay.</p> <p>You might be a good fit if you:</p> <ul> <li><strong>Have a bias toward implementation and shipping</strong>—you are excited to modify real engines and services, not just prototype in research code.</li> <li>Have strong expertise in at least one of the following, and are excited to collaborate across (and grow into) the others:</li> <ul> <li>Systems‑first profile: Large‑scale inference systems (e.g., SGLang, vLLM, FasterTransformer, TensorRT, custom engines, or similar), GPU performance, distributed serving.</li> <li>RL‑first profile: RL / post‑training for LLMs or large models (e.g., GRPO, RLHF/RLAIF, DPO‑like methods, reward modeling), and using these to train or fine‑tune real models.</li> <li>Model architecture design for Transformers or other large neural nets.</li> <li>Distributed systems / high‑performance computing for ML.</li> </ul> <li>Are comfortable working from algorithms to engines:</li> <ul> <li>Strong coding ability in Python</li> <li>Experience profiling and optimizing performance across GPU, networking, and memory layers.</li> <li>Able to ta