Professional Experience#
Verses Inc.#
Sr. ML Research Engineer - Contractor, Remote | Dec 2023 - Present
Details
- Designed and deployed an edge-ready 3D perception stack for reliable warehouse object detection and depth-aware navigation on Jetson-class hardware.
- Developed an YOLOv10n perception pipeline for warehouse-specific objects using a BlenderProc2-curated synthetic dataset, improving mAP from 0.15 to 0.62 at 640×640 and reducing the sim-to-real gap by retraining the synthetic-initialized model on cleaned, real warehouse data to reach 0.58 mAP
- Optimized the end-to-end detection stack (pre/post-processing + YOLOv10n inference) with TensorRT on Jetson Orin NX and desktop GPUs, achieving 26 ms latency at 640×640 for real-time edge deployment
- Integrated a DepthAnythingv2-based depth estimation module into the TensorRT perception pipeline, cutting depth latency from 1 s to 50 ms and delivering a combined depth + detection stack at 10 FPS (640×640) and 13 FPS (378×378) on Jetson Orin NX
Verses Global Bv#
ML Tech Lead (Project: dAIEdge) - Contractor, Remote | Dec 2023 - Present
Details
- Led development of an Active Inference–based routing agent leveraging the perception pipeline to perform obstacle avoidance, optimizing the planning stack and reducing planning time from 7 minutes to 21 seconds
- Developed a PyMDP-based saccading agent using a Tapo security camera for active visual exploration, enabling person detection and tracking and later forming the basis of a peer-reviewed conference publication
- Designed and maintained a 3D simulation pipeline in NVIDIA Isaac Omniverse to develop and evaluate routing agents in realistic warehouse environments before real-world deployment
Publication: Towards smart and adaptive agents for active sensing on edge devices, D. Vyas, M. De Prado and T. Verbelen, HiPeac 2025 Link
Machine Learning Engineer#
TerraLoupe GmbH - Munich, Germany | Oct 2019 – Apr 2020
Details
- My responsibility involved improving toolchains for deep learning experiments to enhance their reproducibility and
trackability.
- Sourced, Cleaned and Automated MLOps pipeline for large-scale aerial image segmentation dataset
- Integrated ML experimentation tracking using Sacred and Omniboard
- Experimented with Deeplab(Semantic Segmentation Model) migration from GPUs to TPUs
ADAS Engineer#
KPIT Technologies GmbH - Munich, Germany | Jan 2019 – Aug 2019
Details
- My task was to create a Vehicle State Monitor that handles high-frequency data from multiple sensors on the car
• Developed a GUI dashboard that can process processes high-frequency data to monitor and visualise the real-time health
of vehicles
- Designed the testing module for the Test Driven Development(TDD) of the Vehicle State Monitor (VSM)
- Contributed to coding standards, code reviews, and source control management
Software Engineer#
CNRS (XLIM Lab) - Poitiers, France | Jan 2017 – Aug 2018
Details
- My responsibility was to integrate and optimize an algorithm for a simulator that visualizes radio wave propagation.
- Optimized run-times by approx. 30% and streamlining system reliability
Publication: CupCarbon: A new platform for the design, simulation and 2D/3D visualization of radio propagation and interferences in IoT networks
Education#
Technical University of Munich | Focus: ML and CV | Oct 2019 – Aug 2023
Bachelors of Technology in Computer Science#
The LNM Institue of Information Technology, B.Tech Computer Science | Jul 2012 – Jun 2016
Skills#
Languages
Libraries#
Libraries
- PyTorch
- Tensorflow
- ROS2
- PyMDP
- Jax
- OpenCV
- Numpy/CuPy