I engineer intelligent systems from sensor to model.
I build end-to-end systems that connect custom sensing hardware, computer vision, machine learning, and reliable data infrastructure.
I am a PhD Fellow at Aarhus University working across intelligent hardware, AI and perception, and ML infrastructure. I take systems through the full engineering lifecycle: measurement design, hardware integration, acquisition software, data pipelines, model training, and quantitative validation.
My work spans automated imaging, bioacoustic sensing, computer vision, distributed GPU training, robotics, and geometric sensor calibration. I am interested in building robotics, intelligent-device, computer-vision, and applied-ML products where hardware and software must work reliably together.
- Custom sensing and automated experimental platforms
- Camera, illumination, and environmental-control systems
- Contact microphones, LiDAR, and time-of-flight sensors
- ROS, Gazebo, UR5, embedded acquisition, and device orchestration
- Computer vision and bioacoustic classification
- Self-supervised representation learning and transformers
- DINOv2, diffusion models, hierarchical VQ-VAE, and NeRF
- PyTorch, OpenCV, model benchmarking, and error analysis
- Python and C++ engineering
- DeepSpeed ZeRO-3 and multi-node distributed training
- Docker, experiment-control software, and metadata logging
- Image acquisition, data ingestion, quality control, and dataset curation
- Computer-vision benchmarking: trained DINOv2 representations on 1.5 million multi-channel images, increasing gene-classification AUROC by 18% relative to an ImageNet-pretrained baseline.
- Distributed-training optimization: implemented DeepSpeed ZeRO-3 across 64 A100 GPUs, reducing epoch latency from 6.4 hours to 1.7 hours, improving throughput by 3.8×, and lowering compute cost by 54%.
- Bioacoustic classification: integrated contact-microphone acquisition, signal preprocessing, self-supervised feature extraction, and classification to reach 96% precision on in-plant pest detection.
- Extrinsic sensor calibration: automated VL53L3CX and VL6180X calibration with ROS, Gazebo, and UR5, validating sensor poses with sub-2 mm 3D reconstruction residuals.
- Automated imaging: built a repeatable camera-control, illumination, metadata, and ingestion pipeline that produced 5,000+ curated images across species.
| Project | Engineering focus | Result |
|---|---|---|
| Automated Experimental Platform | Device control, continuous video acquisition, environmental control, telemetry, and metadata | Modular platform for repeatable long-duration operation |
| Scalable Visual Representation Learning | DINOv2, diffusion, VQ-VAE, DeepSpeed ZeRO-3, and 64-GPU training | 18% higher AUROC, 3.8× higher throughput, and 54% lower compute cost |
| Bioacoustic Pest Detection | Contact microphones, self-supervised learning, signal processing, and classification | 96% precision on faint in-plant pest signals |
| Robot-Assisted Sensor Calibration | ROS, Gazebo, UR5, time-of-flight sensing, and 3D reconstruction | Sub-2 mm reconstruction residuals |
| Computer-Vision Data Pipeline | Camera control, standardized illumination, quality control, and data ingestion | 5,000+ curated images across species |
Python · C++ · PyTorch · OpenCV · DINOv2 · DeepSpeed ·
ROS · Gazebo · Docker · LiDAR · NeRF · Distributed training
- Geometric Calibration of Single-Pixel Distance Sensors — IEEE Robotics and Automation Letters
- Eavesdropping on Herbivores: Using Contact Microphones to Quantify Plant-Insect Interactions — preprint
- Website: dashingzombie.github.io
- LinkedIn: linkedin.com/in/dev-mehrotra
- Email: devd@qgg.au.dk
- Aarhus University: person.au.dk/devd@qgg.au.dk
Building reliable systems where hardware meets AI.