Computer Vision · Robotics · Edge AI

Nihal Soans

Building systems that enable robots to perceive their environment and take intelligent action

Computer Vision Engineer with deep expertise in perception pipelines, stereo geometry, sensor calibration, and real-time deployment on edge hardware. Based in Austin, TX.

"Research should be designed to have practical applications in real-world scenarios"
Experience

Where I've worked

Deploying perception systems in production environments across rail, retail, and industrial robotics.

MLOps & Computer Vision Engineer
BNSF Railway
2025 - Present
  • Developed scalable Python-based deep learning inference pipeline for real-time theft detection on edge devices across BNSF's railway network
  • Achieved a massive boost in real-time model performance through advanced optimization techniques (quantization, pruning, batching)
  • Built an in-house web inference API delivering 60% speed improvements for environments where edge hardware could not run deep learning directly
Python MLOps Model Optimization Perception Pipelines Triton Server TensorRT
ML Engineer — Computer Vision
Radar
2022 - 2025
  • Led development of a 360° camera prototype enabling digital twin creation and spatial analytics across retail environments
  • Integrated scalable C++ CV architectures with LiDAR mapping and deep learning on edge accelerators; delivered real-time performance at ≤30 fps
  • Reduced live streaming bandwidth by 45X through custom compression and frame-drop strategies
  • Achieved 80%+ hardware cost savings via systematic vendor evaluation without compromising system reliability
  • Implemented IMU + LiDAR localization and AI-driven spatial tracking pipelines for real-time occupancy analysis
C++ Digital Twin NVIDIA Jetson LiDAR IMU Computer Vision Physical AI
Robotics & Vision Engineer
SoftWear Automation
2019 - 2022
  • Engineered CV pipelines for textile automation — achieved >99% accuracy in fabric side detection and 90% improvement in seam flip identification
  • Built ROS-based C++ middleware for high-throughput multicamera image acquisition across diverse GenICam interfaces
  • Implemented precision machine vision algorithms via HALCON's C++ API for fabric stretch analysis at 93% accuracy, meeting client specifications
  • Developed a bidirectional C++/ROS framework for synchronized PLC-robot and vision system communication
  • Reduced gantry robot table-mapping time from 1 day to 30 minutes via a novel deep learning interpolation method with Beckhoff PLC + EtherCAT integration
Robotics ROS C++ HALCON GenICam EtherCAT TensorFlow
Skills

Technical expertise

Tools and technologies I use to build perception, robotics, and edge AI systems.

Languages
Python C++
Vision & ML
OpenCV Pytorch / TensorFlow Deep Learning (VLA, YOLO, ReID) Multi Camera Triangulation Camera Calibration Vision Geometry
Robotics & Sensors
ROS LiDAR IMU GenICam EtherCAT / PLC Le-Kiwi
Edge & Deployment
NVIDIA Jetson Triton Server TensorRT ARM Model Optimization MLOps
Research

Publications

Peer-reviewed work in robotics, deep learning, and computer vision.

87
Total Citations
2
h-index
2
i10-index
N Soans, E Asali, Y Hong, P Doshi
IEEE International Conference on Robotics and Automation (ICRA), 2020 · pp. 2153-2159
Deep LearningRoboticsRGB-DImitation Learning
38
Citations
R Achar, PS Thilagam, N Soans, PV Vikyath, S Rao, AM Vijeth
Annual IEEE India Conference (INDICON), 2013
Cloud ComputingLoad BalancingVirtualization
47
Citations
View full profile on Google Scholar →
Projects

Featured work

Selected projects from my personal time.

Multi-Agent Board & Card Games

Models board and card games as Markov Games to find optimal strategies using the Minimax-Q reinforcement learning algorithm. Handles multiple goal states and probabilistic card draws with PyGame visualization.

PythonReinforcement LearningMinimax-QPyGame

Roller Coaster Track Optimizer

Optimizes roller coaster track design using a Genetic Algorithm, balancing excitement variables (G-force, speed, loop count) against safety constraints across age-based demographic groups.

JavaGenetic AlgorithmEvolutionary Computation

Bag to Depth

Python and C++ tools to extract synchronized depth and RGB image frames from ROS bag files. Outputs JPG and NumPy formats for downstream ML pipelines, built on OpenCV and CVBridge.

PythonC++ROSOpenCVDepth Sensing

Google Landmark Retrieval

Kaggle competition entry for large-scale image retrieval of landmark photos. Used VGG16 for feature extraction and K-means clustering for nearest-neighbor matching. Ranked 59 out of 132 teams.

PythonKerasVGG16K-meansKaggle

Automatic Socks Picker Upper

Automated in house robot picks up socks using end to end VLA system and 2D Cameras

le-kiwipytorchPi0VSLAMrobotics

Soans Photography

A hobby — capturing moments through the lens. Follow along on Instagram.

HobbyPhotographyInstagram
Get in touch

Let's build something together

Open to collaborations, consulting, and technical roles in
robotics and computer vision.