Reactor is building the inference platform for real-time world models. Our goal is to democratize access to real-time intelligence to unlock next-generation use cases in media, physical AI and omni-directional models. We already run the largest and fastest catalog of world models on the market, and work closely with all major model labs to run and optimize their world models, including Alibaba, Amazon, and NVIDIA. Our team is made of alumni from leading firms such as Apple, Meta, Amazon, Microsoft, Luma AI, Adobe, Stability AI, and more. We've raised $74M led by Lightspeed, with Amplify Partners, WNDR, NVIDIA, Sapphire, and other Tier 1 investors participating. Jeffrey Katzenberg, co-founder of DreamWorks Animation, sits on our board of directors.
The Role
This is a hands-on role with a lot of scope. You'll help set the technical direction of Reactor's robotics work from the ground up. Day to day, you'll get customer models running reliably on our platform under strict latency budgets, wired into real robots.
We're looking for a generalist across ML, systems and robotics. You understand how a policy behaves at inference time, how batching, scheduling and the network path add to its end-to-end latency, and how its output tensor becomes an action in a robot's control loop.
- Serve robot policies (VLAs, IL/RL models) on GPUs under robot-grade latency budgets, and profile the full path from input tensor to output action
- Build the integration layer between Reactor and robot stacks: ROS/ROS2 nodes and bridges, transport, and low-latency channels carrying observations up and actions down
- Use world models for policy evaluation, and tune how we serve rollouts for throughput, fidelity and turnaround time
- Bring teleoperation and human-in-the-loop experience to our real-time transport: synchronization, jitter and failover for a live control loop
- Debug with customer robotics teams across the whole loop: sensor capture, network, inference, action output, robot execution
- Help shape the technical roadmap and set the engineering bar for the team around you
- You've run learned policies or ML models on real robots or autonomous systems, and debugged failures from perception through policy, control and hardware
- Strong ML understanding. You know how modern robot policies (VLAs, imitation learning, RL) and world models are built (transformer backbones, action tokenization or diffusion heads, observation encoders) and how they behave at inference time. You won't be training them.
- Depth on inference systems: batching and scheduling, GPU utilization, p99 latency vs. throughput, and running models in real time against physical systems
- Strong C++ and Python in production, and comfortable changing performance-critical code
- Working experience with ROS/ROS2, DDS and robot middleware, from message definitions to transport QoS settings
- Range across ML, systems, cloud infra, networking and robotics. You'd rather fix whatever's blocking the robot than stay in one layer.
- The judgment to lead: setting direction, making tradeoffs explicit and representing Reactor in front of customer engineering teams
Bonus Points
- World models, video or diffusion models, or learned simulators for policy evaluation or training
- Teleoperation: low-latency control loops, sensor fusion for human-in-the-loop tools, demonstration data collection
- Inference optimization: INT8/FP8 quantization, CUDA/Triton kernels, TensorRT, vLLM/SGLang-style serving
- Real-time networking: WebRTC, RTP/SRTP, QUIC, jitter buffers, frame budgets
- Infrastructure: multi-tenant GPU serving, autoscaling, data pipelines, eval tooling
- Systems performance: Linux internals, perf/eBPF/GPU profilers, PREEMPT_RT, cgroups, camera pipelines
- Humanoid or dexterous manipulation
- Early-stage startup or founding-engineer experience
Logistics
We're in person in San Francisco
- Competitive SF salary and meaningful early equity
- Visa sponsorship and relocation support
- Generous health, dental, and vision coverage