Physical AI engineer. I build perception, reasoning, and autonomy systems for humanoid robots and autonomous vehicles. I like figuring out what a system actually needs versus what it says it needs.
MoRAL - VLM fine-tuned for 3D autonomous driving scene reasoning on nuScenes. BEV + radar inputs, zone F1=0.89, 42 tok/s at 4.61 GB VRAM. Fine-tuned 2B outperforms zero-shot 8B on safety-critical decisions. IEEE IMC 2026 under review
Toborlife - Real-time multimodal data pipelines for Unitree G1 humanoids and Go2 quadrupeds. MCAP/zstd, Parquet, LiveKit WebRTC teleoperation, Jetson Thor/AGX fleet feeding VLA and RL training.
SJSU IAS Lab - Sensor fusion combining LiDAR and stereo cameras, PointPillars 3D detection, ROS2 nav stack, SLAM, NVIDIA Warp GPU acceleration across 50+ Jetson Orin/Nano devices.
Robotics: ROS2, Isaac Sim, Isaac Lab, MuJoCo, Gazebo, SLAM, sensor fusion, sim-to-real
AI / ML: PyTorch, LoRA fine-tuning, VLMs, vLLM, CUDA, NVIDIA Warp
Perception: LiDAR, stereo cameras, radar, BEV, PointPillars, nuScenes, ICP/NDT
Languages: Python, C++, Go, TypeScript, Bash, SQL
3 years production SWE at SmartHub.ai: 10K+ cameras, 15K+ devices across enterprise clients, 97% OTA pipeline improvement (5 hours to 7 minutes), Python and Golang SDKs adopted by 4K+ devices, YOLO models on live security feeds.
🏀 Basketball 🚵🏽♂️ Cycling 🥾 Trekking 🏍️ Long drives 🐲 D&D 📚 Books 🎥 Movies
Robotics SWE · ML/AI Engineer · AI Agent Systems · SWE | US-based, open to relocation, May 2026
🌐 ambarishgk.me |
🤗 HuggingFace |
💼 LinkedIn |