Founding Researcher (Robotics)

Noida

Responded to 90% candidates

Location: Noida, Delhi NCR

Starting Date: April 2026 (or sooner)

Compensation: 50 LPA + Significant ESOPS

As a founding engineer (Robotics) you will build the data and model infrastructure powering our business - from raw video ingestion through task annotation to model-ready formatting. This is a high-context, high-impact role at the frontier of embodied AI, ideal for someone who moves fluidly between research and production. You won’t be handed a roadmap; you will write it.

WHAT YOU WILL BUILD

Egocentric Data Pipeline: Own end-to-end: raw video ingestion → task annotation → model-ready formatting. Quality, throughput, and scalability are all yours
LLM-Powered Annotation: Design and maintain workflows that translate raw egocentric footage into structured, high-fidelity labels at scale, with minimal human bottlenecks
VLA Integration: Stay at the frontier of VLA research (SmolVLA, π0.5, and what comes next). Rapidly integrate advances into our stack before they are mainstream
Lab Interface: Serve as the technical bridge to external AI labs and research partners navigating data format requirements, aligning on evaluation protocols, and maintaining trust.
Model Optimisation: Profile and optimise CV models (e.g. HaMeR) end-to-end, from algorithmic improvements through TensorRT/Triton deployment for high-throughput GPU execution.

WHAT WE NEED

Experience: 3+ years in applied AI/CV engineering, robotics research, or a closely related field with demonstrably exceptional ability
VLA Depth: Deep, current knowledge of OSS VLA architectures and the embodied AI landscape. You read the papers the week they drop.
PyTorch Mastery: Expert-level: model training, custom layer design, systematic profiling. You know what’s happening at the CUDA kernel level.
Inference Engineering: Proven deployment and optimisation with TensorRT/Triton for high-throughput production GPU execution
Annotation at Scale: Demonstrated experience building LLM-assisted video annotation pipelines - not toy demos, real throughput.
Research Communication: You can hold a precise technical conversation with a frontier lab and leave with exactly what both sides need

WHY THIS ROLE

A front-row seat to building the most important dataset for physical AI from the ground up
Work alongside a small, high-calibre team from UC Berkeley, IIT Bombay, and BITS Pilani
Direct, day to day, collaboration with frontier research labs in AI & robotics
Founding equity and the chance to define how embodied AI data infrastructure is built, not just participate in it

Name *

Email *

Walk us through a video data pipeline you have built or substantially owned in production *

What was the throughput target? What broke first, and how did you fix it? We want the unglamorous version

Pick a VLA paper published in the last 12 months that you think is underrated. Make the case for it *

We are testing depth of engagement with the literature - not whether you picked the “right” paper

You have a CV model in production that is too slow. Walk us through your profiling and optimisation process from diagnosis to deployment *

Be specific: tools, bottlenecks you have actually encountered, and the trade-offs you made

Describe a time you had to design a data format or annotation schema from scratch for a research partner. What questions did you ask, and what did you get wrong the first time? *

We are looking for technical humility and communication instincts - not perfection

What is the hardest thing about making egocentric video data useful for training embodied AI models? What is your current thinking on how to solve it? *

There is no established playbook here - we want to hear your original view

Describe a time you used an LLM to automate a task that would traditionally require human judgment. How did you validate it was working? Focus on the validation strategy - that’s the hard part *

Focus on the validation strategy — that’s the hard part

What is the most interesting failure you have had in an AI/ML system you built? What did it teach you about the gap between research and production? *