Founding Software Engineer, Robot Learning
Role summary
Kovari is seeking a Founding Software Engineer to own the robot perception stack end-to-end. This role involves researching and developing high-reliability manipulation policies for deployed robots in real-world hospitality environments. You will work with multimodal sensor data, debug complex policy failures, and optimize for real-time inference on edge hardware. The ideal candidate has experience deploying robot policies, working with multimodal inputs, and optimizing for constrained devices using tools like TensorRT or ONNX Runtime. This is an onsite, full-time role in San Francisco, requiring a strong commitment to the mission and rapid iteration.
Company Description
At Kovari, we're rethinking how physical work gets done in the age of robotics. We believe building robots that can move the economy is one of the most important endeavors in technology.
Our first goal is to build general-purpose robots for hospitality to take on physical, repetitive work that keeps the hospitality industry operating. The last mile problem for proliferating useful robots into businesses is a first class innovation problem itself. We aim to marry deep commercial understanding with fast paced innovation to create robots that move the industry. Since inception, we have raised over $6M to carry out our mission from industry leading investors.
We are obsessed with rapid iteration, engineering rigor, and deploying real machines into real environments. The next decade will compress a century of progress in robotics, and we're looking for people who want to leave their fingerprints on that future.
We are based in San Francisco and work in-person.
The Role
You will own Kovari's perception stack end-to-end—from raw sensor data to actionable representations for both learned policies and classical control. Your systems will run on deployed robots in real hotel environments, handling the messy realities of variable lighting, glass surfaces, temporary obstacles, and repetitive architecture.
What You'll Do
- Research and develop high-reliability manipulation policies designed for high-velocity deployment and iteration
- Operate in a fast data flywheel across multiple data modalities
- Deep debug failure modes in transformer and diffusion policy field deployments
- Optimize policies for real-time (~10hz) inference on edge hardware
What you bring
- Experience deploying robot policies on hardware No preference between model-based learning, reinforcement learning, or imitation learning
- Sim-to-real or real robot data
- Experience building policies with multimodal inputs (vision, depth, force/torque, proprioception)
- Experience with deep optimizations for constrained edge devices TensorRT, ONNX Runtime, or TVM for inference optimization
- CUDA kernel optimization
- Ideally, contributions at major robotics/ML conferences (CoRL, RSS, ICRA, NeurIPS)
Values
- Pace of learning trumps everything else.
- Refining our craft is something we pursue relentlessly.
- Low ego, high ownership.
- Commitment to the mission. We work in-person, and this isn't a 9-to-5. We're building something hard, and we need people who are all-in.