Martin Weyssow

- Led end-to-end development of software vulnerability detection ML systems on a large-scale research project in collaboration with GovTech Singapore, focused on developing and deploying LLM-based detection models in CI/CD production pipelines. - Designed and implemented post-training pipelines combining RLHF with parameter-efficient fine-tuning on small language models (<7B parameters), achieving ~70% training and inference cost reduction while boosting detection accuracy against larger and commercial models by more than 35%. Stack: PyTorch, HuggingFace (Transformers/PEFT/TRL), vLLM, DeepSpeed. - Designed and deployed end-to-end vulnerability detection system integrated with GitLab CI/CD pipelines capable of processing thousands of commits daily while generating structured reports to help developers with vulnerability mitigation. - Engineered synthetic data generation framework using multi-agent systems (LangGraph, OpenAI/Anthropic APIs) to create diverse vulnerability datasets and improve coverage of common vulnerability categories (CWEs). - Implemented model calibration techniques at inference to reduce false positives and meet deployment requirements, e.g., low noise signal for developers. - Mentored cross-functional teams of 15 students and engineers, delivering research prototypes that transition to production systems.

- Designed and developed deep learning MOOC for Desjardins employees, covering generative modeling, computer vision, reinforcement learning, and graph neural networks with theoretical concepts and practical applications. - Created structured learning materials and hands-on examples to introduce deep learning concepts to non-technical audiences across the organisation.

- Developed and implemented parameter-efficient fine-tuning techniques (LoRA, quantization) reducing LLM training costs by 10x while maintaining code generation performance on benchmark tasks (HumanEval, MBPP). - Built continual learning pipelines using replay buffers and regularization techniques enabling language models to adapt to evolving software ecosystems (new frameworks and third-party libraries) without catastrophic forgetting. - Established international research partnership between University of Montreal and SMU, securing $100K in NSERC competitive funding to support cross-institutional ML research.