Welcome!
I’m Mohammadreza Rostam, a PhD candidate at the Control Engineering Laboratory within the University of British Columbia. Simultaneously, I’m a senior deep learning researcher and developer at Picovoice. My expertise covers:
- Crafting deep learning models: RNNs, CNNs, Transformers, and GANs.
- Proficient in CUDA programming for GPU optimization
- Implementing NLP tasks on resource-limited devices.
- Time-series forecasting techniques.
- Optimization and Model Predictive Control.
- Various aspects of control theory: adaptive, robust, nonlinear, and optimal methods.
- Creating data-driven models.
- Developing firmwares for embedded systems
Current Work
As a senior deep learning researcher at Picovoice, I’ve led the development of advanced speaker recognition (Eagle) and diarization (Falcon) engines from start to finish. I also integrated speaker diarization into our speech-to-text engine, Leopard, with minimal impact on performance.
Currently, I’m focused on pioneering new neural network architectures for Natural Language Processing (NLP) applications and optimizing core runtime engines for speed and efficiency across different platforms. I’ve recently designed a lightweight, high-performance inference engine specifically for LLMs.
In the past, I played a key role in expanding language support for our Speech-to-Text engines, ensuring its effectiveness across diverse linguistic contexts. My experience also includes working with various SDKs and technologies like Rust, Node.js, and WebAssembly.
Overall, my role at Picovoice involves driving innovation in deep learning research, fine-tuning runtime engines, and expanding the capabilities of our speech technologies to meet diverse application needs.
Current Research
My area of expertise and research focus is on creating smart control algorithms that enhance the efficiency of mechatronic systems. I am presently conducting doctoral research on a novel control system that merges the model predictive control approach with Gaussian process machine learning technology. This cutting-edge system is intended to handle unknown disturbances with quasi-periodic patterns, such as those associated with renewable energy systems’ energy requirements.
Previous Work
Before starting my PhD, I gained experience as an embedded software/mechatronic engineer in multiple small startups. During this time, I sharpened my programming skills for Bare-Metal, RTOS, and Linux systems, focusing on firmware development and working extensively with various sensor types.