Vancouver, Canada
I'm Uroš, a computer science student working on LLM research and embedded systems. My work ranges from LLM evaluation and unlearning research to STM32/ESP32 firmware, sensors, BLE, and edge ML.
Currently building
Selected work
Wearable movement analysis and real-time exercise feedback
Streams 6-axis IMU data from an ESP32-C5 at 50 Hz, learns a mounting-independent motion axis with PCA, segments repetitions, and delivers bounded, fail-soft coaching through a live dashboard and voice pipeline.
Python · FastAPI · React · TypeScript · ESP32 · BLE · Signal processing
Gesture-controlled embedded lock prototype
An STM32F411 classifies motion locally from MPU6050 IMU data, then sends accepted gestures over UART and BLE through two ESP32-C5 boards to actuate a servo lock.
C · Python · STM32 · ESP32 · TinyML · BLE · I2C · UART
Decoder-only transformer implementation
Built a complete transformer with causal multi-head attention, pre-LayerNorm residual blocks, autoregressive generation, and configurable sampling strategies.
Python · PyTorch · NumPy · Machine learning
Small automatic differentiation framework
A NumPy-based autograd engine with tensor operations, broadcasting, neural-network layers, optimizers, gradient clipping, and tests for gradient correctness.
Python · NumPy · Autograd · Neural networks
Undergraduate research at Simon Fraser University
Extended the RWKU framework with an entity-aware benchmark and evaluated 54,000 responses across three LLMs, measuring target forgetting while preserving similarly named entities.
Python · LLM APIs · BM25 · ROUGE-L · Data pipelines