Diego Pousa

I run Train Them Al, a boutique data ops studio for robotics teams that need policy-ready VLA datasets without the $500K minimums and 6-week timelines of enterprise vendors.What we do:• Egocentric data collection — distributed operator network across LatAm and the Philippines, capturing first-person video of household, kitchen, and workspace activities at scale.• Teleoperation sessions — portable Quest 3 + GELLO setups, on-site or remote, on your hardware (ALOHA, UR, Franka, custom rigs).• VLA action annotation — episode-level language instructions, subtask segmentation, success/failure labels, contact tags. OpenVLA · pi0 Octo ACT.Every delivery includes a policy smoke test before we ship — we run a mini training pass on 10% of the dataset and fix the data if the policy doesn't converge. No other vendor in our segment does this.Background: I develop automated action labeling Al at a venture-backed Al company. Train Them Al applies the same techniques — fine-grained temporal segmentation, automated QC scoring, action vector normalization — to teams that need data ops as a service.Pilots start at 100 demos, ship in 2 weeks, delivered in LeRobot, RLDS, HDF5, or robomimic.