The model stack
Kriya, Prana and Karma — one policy family, one world model and one omni-conditioned foundation model, learning across manipulators, humanoids, quadrupeds and drones.
Careers
Machines are moving into the world. Someone has to make that safe.
We build the perception, action and safety models that let them — and we ship them to 500,000 robots in 140 countries, not to a paper.
Three problems, one stack.
Whichever team you join, you are working on a model that has to survive contact with the physical world.
The safety layer
Every action verified before it reaches an actuator. Severity over keywords, speech and vision judged together, and the same semantics carried into embodied motion.
The flywheel
Deployed fleets return evidence, evidence becomes training signal, and the next model goes back out. The loop is the product.
Open roles
Mumbai, on-site. If none of these is quite you but the problem is, write anyway — we read every one.
Research Scientist, Robot Policy Learning
Train and evaluate the Kriya policy family — imitation and reinforcement learning across manipulation and locomotion, on real hardware rather than only in simulation.
RL / imitation learning · Manipulation · Sim-to-real
ApplyResearch Engineer, World Models
Build Prana, the world action model: learned dynamics that let a robot predict the consequence of an action before it commits to it.
World models · Video prediction · Large-scale training
ApplySafety Researcher, Multimodal Evaluation
Own how we measure harm. Build the benchmarks, red-team the model across speech, image and video, and turn escalation patterns into policies the runtime can enforce.
Evaluation · Red-teaming · Trust & safety
ApplyEmbodied Safety Engineer
Carry verification semantics into physical action — pre-execution checks on force, velocity, reach and occupancy, across manipulators, humanoids, AMRs and drones.
Robotics systems · Real-time · Controls
ApplyInfrastructure Engineer, Training & Data
Run the pipelines behind the flywheel: ingest from deployed fleets, curate and label at scale, and keep large multi-node training jobs fed and reproducible.
Distributed training · Data pipelines · GPU infrastructure
ApplyApplied Scientist, Speech & Audio
Extend Shruti — recognition, paralinguistics and fluency scoring for children's speech, in classrooms and in homes across 140+ countries.
ASR · Paralinguistics · On-device
ApplyHow we hire
Four steps, about three weeks, no take-home that outlasts a weekend.
01
You write
A note about what you have built and which of these problems you want. No cover letter.
02
We talk
Forty-five minutes on your work and ours, with the person you would work beside.
03
You build
One focused problem from our actual queue.
04
You decide
You meet the team, we make an offer, and you ask us anything still unanswered.