Artificial Intelligence Intern

About the job
We record people teleoperating a robot arm through pick-and-place tasks and train policies on that video. Between recording and training sits a pipeline that nobody sees and everybody depends on — it filters out blurred, blown-out, occluded, or frozen frames, locates the end-effector, crops around it, and produces clean training clips with a label showing how much of each recording survived.
That pipeline is this internship. It's unglamorous and it's the highest-leverage part of the project: a policy trained on a frozen stream doesn't fail loudly — it quietly learns that the world doesn't move when you act on it.
What you'll do
Own the QC stage — sharpness, exposure, occlusion, near-duplicate, and region-of-interest checks on raw frames, with thresholds you can defend from the data
Turn raw recordings into training clips: extraction, cropping, resizing, and keeping the crop stable across a clip rather than letting it jitter
Decide which recordings are unusable and say so — a rejected clip is worth more than one quietly passed
Use a VLM for checks a metric can't express (did the grasp happen, does the scene match its task label) — and know when it's the wrong tool
Audit labels, sample by hand, and measure your own pipeline against what you find
Work with Claude daily, and know exactly where you stop trusting it
What we're looking for
Hands-on comfort with OpenCV, PIL, and numpy on your own images — not just benchmark data
You look at the distribution before picking a threshold
You've called a vision-language model yourself and have a sense of what it costs
You check things — if a script says 53 frames passed, you open a few and look
You can say what you didn't verify, without being asked twice
Final-year student or recent graduate, able to work full-time from Bengaluru for 6 months
Nice to have (genuinely not required)
Robot learning or imitation-learning datasets (LeRobot, Open X-Embodiment) · FFmpeg · off-the-shelf detection/segmentation on your own data · annotation QA · large-scale dataset tooling · Kaggle, open-source work, or write-ups
Why this is worth your time
You'll own one stage of a real pipeline end to end and watch it decide what a robot learns. It's a small team, so nothing you build sits behind three layers of approval — and data quality is a skill that transfers to every ML role you'll go on to hold.

Work arrangement
No