@iceman404

John Subba

Systems & Heterogeneous Compute Engineer

I build real-time systems close to the metal: CUDA/HIP kernels, GPU-accelerated computer vision for AR/VR, and embedded sensor fusion. Currently researching GPU architectures: benchmarking memory hierarchies and warp-level primitives across NVIDIA and AMD.

Live CV Demo View Projects
<15ms Real-time inference on mobile GPUs
30% Memory bandwidth saved in AR/VR pipelines
<10ms End-to-end sensor-fusion latency
3+ Production AI systems architected & mentored
Portrait of John Subba (iceman404)

John Subba

GPU & Computer Vision Engineer · Kathmandu, Nepal · B.IT, Purbanchal University

Languages: C++17/20, C, Python, SQL, x86/8052 Assembly

Heterogeneous Compute: CUDA, HIP/ROCm, Metal Shading Language, OpenGL ES 3.1/3.2, Vulkan, WebAssembly

Computer Vision & ML: OpenCV, MediaPipe, PyTorch, TensorFlow, scikit-learn, Hugging Face Transformers

Systems: Linux kernel modules & syscalls, Windows kernel drivers (WinDbg), POSIX threads, IPC

Robotics & Embedded: ROS2, sensor fusion (KF/EKF/UKF, SO(3)/SE(3)), ESP32, 8052, Raspberry Pi, LoRa

Tooling: Git/GitHub Actions, Docker, CMake, Clang-Tidy/PVS-Studio, ASan/UBSan/TSan

Independent GPU architecture research: porting CUDA vision kernels to AMD ROCm (HIP) and benchmarking performance, memory bandwidth, and thread occupancy across NVIDIA and AMD hardware.

Spoken Languages: English, Nepali, Hindi

Education: Bachelor's in Information Technology (Purbanchal University, 2021 to 2025)

Beyond the terminal: Open-source UAVs, quadcopter flight tuning, and a Rubik's Cube that never survives past two minutes.

Project Showcase

Selected work across GPU systems, computer vision, embedded hardware, and backend infrastructure, each with a full write-up.

MIRA DTMF appliance automation hardware

MIRA

Dial-in home automation: DTMF tones from any phone are decoded by an 8052 MCU and toggled into relay commands, with IR sensors and SMS alerts.

Embedded C · 8052 Assembly · SIM800L

View Details Video GitHub
LoRa geofencing tracker hardware

LoRa Geofencing

ESP32 + LoRa RA02 (433 MHz) GPS tracker with geofence alerting: long-range, low-power boundary monitoring for UAVs and remote assets.

ESP32 · LoRa · GPS · Embedded C

View Details GitHub
VisionGuardianCommander surveillance dashboard

VisionGuardian Commander

JavaFX MVC surveillance suite: real-time OpenCV face detection with email alerts and MySQL-backed event logging.

Java · JavaFX · OpenCV · MySQL

View Details Video GitHub
FaceMood music player interface

FaceMood Music Player

Mood-aware desktop player: MediaPipe reads your facial expression in real time and the queue follows, with cross-platform audio fallbacks.

Python · MediaPipe · OpenCV · PyQt6 · Spotify API

View Details GitHub
Face counter system with ESP32 hardware

Face Counter System

Distributed ESP32-CAM edge-ML system: streams video, counts faces on-device, and triggers alerts over a live server dashboard.

ESP32 · Edge ML · Python Server

View Details GitHub
Fine-tuning with ResNet50 training visuals

Fine-Tuning with ResNet50

Transfer learning for flower classification: 88% accuracy via augmentation and head replacement, without significant overfitting.

Python · PyTorch · ResNet50 · Transfer Learning

View Details GitHub
Advanced Redis clone server rack visual

Advanced Redis Clone

High-performance in-memory key-value store in modern C++: RESP protocol, event-driven I/O, caching, and replication for distributed setups.

C++17/20 · Networking · Systems Design

View Details
Open hardware ESP32 drone in flight

Open Hardware Drone

Fully customizable quadcopter on a 330 frame: 4× BLDC motors, 8-inch props, ESP32 telemetry, and hands-on PID flight tuning.

ESP32 · BLDC · Flight Control · Hardware

View Details
× Enlarged project screenshot