Experience & Projects
-
Jul 2025 – Sep 2025
Software Engineer — Techex
Worked on performance optimisation and streaming infrastructure at Techex.
- Refactored React components to avoid unmount/remount cycles, optimised MobX usage, and stabilised the UI
- Implemented process-specific CPU usage tracking and real-time charts for module statistics
- Developed WHIP/WebRTC Rust output module to parse incoming PES packets and stream RTP packets, debugging with Wireshark
- Automated encoder video quality testing (VMAF), saving over a week of manual effort per test round
-
Feb 2025 – May 2025
Full-stack Car Inventory & Invoicing System
Built a complete vehicle inventory and invoicing system.
- FastAPI backend with SQLite database
- Custom JS/HTML/CSS frontend with responsive layout
- Automated data ingestion using a Selenium scraper with cron scheduling
- PDF invoice generation with digital signing for secure payment tracking
-
Aug 2024 – Ongoing
Self-hosted Server
Self-hosting multiple services and applications on a Raspberry Pi Zero W.
- Hosting portfolio, Minesweeper, and personal todo-list web apps
- Configured Nginx reverse proxy, Cloudflare DDNS via API, and CI/CD with GitHub Actions
- Implemented SSH access with keys + fail2ban for security
- Automated off-site snapshot backups for resilience
- Monitoring with Prometheus + Grafana for remote alerts
-
Nov 2024 – Dec 2024
Convolutional Neural Network (from scratch in C++)
github.com/madmax755/cnn-from-scratch
Built a CNN library from scratch in C++ for image classification.
- Implemented convolutional, pooling, and dense layers
- Custom Tensor3D class for efficient data handling
- Custom optimisers (SGD, Momentum, Adam, AdamW with gradient clipping)
- Model serialisation (save/load)
Results:
- Achieved 99.4% accuracy on MNIST with augmentation (demo)
-
Sep 2024 – Oct 2024
Feedforward Neural Network (from scratch in C++)
github.com/madmax755/mlp-from-scratch
Implemented a classic MLP neural network entirely in C++ without ML libraries.
- Highly configurable network topologies with multiple activation functions (ReLU, Sigmoid, Softmax)
- Optimisers: SGD, Momentum, NAG, Adam, AdamW
- Multi-threaded training support
- Model persistence and evaluation metrics
Results:
- Reached 91% accuracy on MNIST digit classification