Simran Lalwani.
ML engineer · 3D Perception · San Francisco
Building things that see, sense, and act.
Same aerial frame, four Cosmos-Transfer2.5 outputs. The pipeline that generated and validated ~50k annotations on A100-80GB, feeding controlled domain-transfer ablations for downstream detectors.
view full demorecent
all projects →about
I have an MS in Data Science from RIT and have been building perception systems for autonomous navigation, edge AI, and public safety.
My research spans neuroevolution for drone flight safety (RNN architecture search on real NGAFID telemetry with Prof. Travis Desell) and domain generalization for time-series classification across unseen distributions using ECG data, with Dr. Linwei Wang.
At OWL Autonomous Imaging I built MGNet, a thermal+RGB dual-encoder fusion model for panoptic segmentation and monocular depth estimation, improving segmentation performance 20% over an RGB-only baseline. I also shipped real-time super-resolution and denoising for 16-bit sensor streams and built a multimodal pipeline across 4,000+ hours of thermal, RGB, and LiDAR capture, improving detection mAP 15%. At Alta Potentia I deployed object detection on NVIDIA Jetson at sub-100ms latency with TensorRT, processing 500K+ video frames for retail loss prevention.
Most recently: Cosmos-Overwatch , a synthetic data platform using NVIDIA Cosmos-Transfer2.5 to generate domain-shifted aerial footage (rain, fog, thermal, night-fire) with dual edge+depth control, validated across ~50K MOT-compatible annotations.
If I had to bet on where I go next, it's perception engineering in autonomous vehicles, robotics, or drone systems: 3D perception, multimodal sensor fusion (camera/LiDAR/radar), models running at scale under real-world constraints - though I'm keeping an open mind for the right ML or software role too.