My Professional Journey

• Technical lead and feature owner within the Dose Control cluster.
• Responsible for feature intake, software design, work package creation, testing strategy, documentation, implementation and end-to-end delivery.
• Driving alignment and delivery across multiple stakeholders.
• Successful delivery of multiple complex features with extra scopes under tight deadlines.

Tools: Python, C, C++, Robot Test, GTest, Git, Jira, CI/CD, Jenkins, Docker, Confluence, Polarion, ClearCase, Sparx EA

• Designed and evaluated an ultra-compact object detection pipeline intended to fit under 1 MB for ultra-low-power MCUs such as STM32H, STM32F, STM32L and Arduino Portenta-class devices.
• Targeted low-cost battery-powered supermarket shelf monitoring/SKU systems that wake periodically, capture an image, and detect shelf products with minimal memory and compute budget.
• Worked on SKU110K, Grozi 120-3.2k and Migros-style retail object datasets with large category diversity, where traditional shelf-specific computer vision approaches were not sufficient for scalable deployment.
• Explored and implemented multiple model and training improvements for custom Object Detection Model development inspired by MCUNet, MobileNet's and YOLO-v3 to V11 approaches, including model quantizations, knowledge distillation, loss-function improvements, augmentation ideas, and lightweight deployment-oriented design decisions.
• Researched and benchmarked all state-of-the-art MCU targeted detection models and frameworks such as Edge Impulse, STM32-AI-Zoo, MIT Tiny Engine, NXP eIQ.
• The research concluded after showing that the target accuracy was not achievable within the intended ultra-low-power memory and compute limits.

Tools: C, C++, Python, PyTorch, TensorFlow Lite, TensorFlow Lite Micro, OpenCV, STM32-AI, SKU110K, Grozi, YOLO-v3 to V11, MobileNets

Images from thesis work

← →

Result image 1 Example successful inference / detection output

Result image 2 Example non-successful inference / detection output

Result image 3 Example non-successful inference / detection output

Auto Anchor Selection Grozi Dataset Object W-H Graph for Auto Anchors

• Developed real-time driver monitoring and pupil tracking systems for automotive 3D AR HUD applications.
• Implemented and deployed optimized neural network inference pipelines on several devices such as Qualcomm Snapdragons, NXP, Blaize and Nvidia Xavier platforms.
• Built end-to-end ML workflows to develop object detection models covering dataset generation, training, evaluation, optimization and embedded deployment.
• Integrated neural network models into production automotive systems under strict latency and performance constraints.

Tools: C++, Python, TensorFlow, PyTorch, TensorRT, OpenCV, OpenVINO, OpenVX, OpenGL, SNPE, QNN, Hexagon DSP, Git, Jira, Yocto Linux

Media

← →

Video

Eye Tracking & Rendering How our product projects images to eyes at real-time.

Video

AR HUD Vision What drivers can see at windshield in real-time.

CES 2024 Showcase High engagement around our product at CES 2024.

Test image in windshield Our HUD test image in windshield at real-time.

Example rendered image Rendered speedometer display at our HUD

Tools for development Debugging opto-mechanical image errors at software level.

• Led end-to-end embedded software development for an in-house STM32-based control board used in warehouse AGV / Autonomous Mobile Robot applications.
• Contributed to prototype bring-up and production-ready board revisions, working closely with hardware development across board validation, peripheral integration and system-level testing.
• Developed embedded drivers and low-level firmware for sensors, motor drivers, communication buses, power-control units and real-time control algorithms.
• Implemented board-level diagnostics, debugging support and hardware/software validation workflows to improve reliability during robot integration and field testing.
• Supported embedded device development topics including MCU configuration, communication interfaces, actuator control, power sequencing, firmware debugging and production readiness.
• Handled QR detection/scanning tasks and researched multi-object tracking in unknown environments.

Tools: Python, C, C++, MATLAB, STM32, OpenCV, TensorRT, Jira, Zoho, Git

Images for this role

← →

Video

Robee and friends Robee hanging out with friends at Warehouse

Video

Robee ♥ Robee against obstacles

Video

Robee charging Robee needs to charge

Dev. In-house Board Oscilloscope debugging, verifications

HIL Test HIL Tests with schematic verifications

Computer Vision Research Consecutive frames captured and feature matching represented

• Worked on computer vision R&D including instance segmentation, object detection and skyline detection for UAV sense-and-avoid systems.
• Developed a user-facing application for Teknofest that detected facial landmarks and overlaid pilot suits on participants.

Tools: C++, Python, OpenCV, Keras, PyQt

Images for this role

← →

Video

Semantic Segmentation R&D work on semantic segmentation for UAV Sky Detection systems.

Air Segmentation Semantic segmentation visualization from UAV wing.

Skyline Detection Skyline Detection algorithm with binary line drawing

Skyline Detection again Skyline Detection algorithm with False Positives

• Worked on Battery Management System software, battery balancing and communication protocols over CAN Bus and ModBus.
• Contributed to a drowsiness-alert system based on driver facial expressions.
• Internship work included ATmega-based battery charger device programming with screen, buzzer and potentiometer support.

Tools: C, Python, MATLAB, OpenCV, Assembly, Atmel

Images for this role

← →

Video

ATmega-based Battery Charger Internship project demonstration

AKAZE Features Matching Tracking-by-detection method tried within 2 consecutive frames AKAZE features.