I’m a robotics engineer, and most of my work sits where state estimation, semantic mapping, and autonomous navigation meet real hardware. I like the problems that only show up once a robot leaves the lab: clocks that drift apart, sensors that get knocked out of alignment, and maps that go stale the moment someone moves a pallet.
Right now I’m a Founding Robotics Engineer at Operation Autopilot, where I own the C++/ROS 2 autonomy stack for a mixed fleet of custom AMRs, Unitree Go2 quadrupeds, and a Unitree G1 humanoid. A lot of my time goes into estimation: a GTSAM/iSAM2 factor graph that fuses IMU preintegration, visual tracking, leg kinematics, and wheel odometry, with PTP keeping every sensor on the same clock. On top of that I built AUTO-MAP, a shared 3D metric-semantic map where CLIP, DINOv2, and VLM features live in a scene graph that remembers when things changed, so a robot can answer a plain-language question about a building on its own Jetson Thor. I also set up the unglamorous half that keeps a fleet alive: CI on every PR, rosbag replay regressions, Docker images, and over-the-air deploys with rollback.
Before that I spent about a year at Vinayak Technology on a team building construction AMRs that had to work without GPS. I wrote C++ ROS 2 drivers, calibration, and custom Nav2 planner and controller plugins, and extended LIO-SAM and FAST-LIO2 with NDT prior-map alignment, loop closure, and online extrinsic recalibration for the days a sensor mount took a hit. An Isaac Sim twin with tuned inertia, contact, and actuator models let us debug the robot before driving out to site.
Earlier, at ISRO, I worked on software for a SAR payload data recorder and on Gaganyaan crew-cabin display software in C++ and OpenGL ES. At Samsung R&D I benchmarked detectors and deployed a TensorRT-optimized YOLOv4 on smart-refrigerator edge hardware, which earned an Excellence Award.
During my master’s I kept coming back to two things. One was pedestrian-aware driving on a real Polaris GEM e4: first AutoShield, a TTC-gated yield and stop supervisor, then Adapt, where an 8.18M-parameter diffusion model forecasts pedestrians five seconds out and GPU-accelerated MPPI plans around them. The other was contact-rich manipulation, like a contact-gated residual SAC policy that seats a USB-A plug on a UR5e. These days that curiosity looks like training manipulation policies, trying VLA models on everyday tasks, tinkering with human-robot interaction, and deploying world models on whatever robots I can get running.
I have a master’s in Autonomy and Robotics from the University of Illinois Urbana-Champaign and a B.Tech in Electronics and Communication Engineering from VIT. The full write-ups, films, and code are in projects.
