I work on getting machines to understand how people move, and how they interact: where two people touch, how their limbs wrap, how far apart they really are, measured in 3D and over time from ordinary video. Independent researcher, based in Bangalore. More at dhruva.life.
Research (one repo per thread, numbers sourced, negative results kept; index and timeline)
| engage3d | Two-person engagement in 3D from one camera: pre-registered experiments, a public model |
| moveharness | The harness and scorer from my CAISc 2026 paper on faithful human-movement understanding |
| movement-research | Measuring movement under contact from one camera, retractions kept |
| kakashi-autolabel | Auto-labelling combat video with a VLM agent and SAM 3.1, checked by hand |
| limb-entanglement | Probing a 27B VLM for spatial information, and a tangle tracker |
| fast-slow-vision | Can a small fast model describe people well enough to reason on? |
| vlm-experiments | Where it started: VLM pipelines for long grappling video |
Also built: uru-mon (a creature-collecting game across Bengaluru) · pushit.tv (the indie music scene, one tape at a time) · keepr (a dead man's switch for crypto) · kakashi.ai (an AI training partner for jiu-jitsu)