Md. Sazid Uddin
Research vision
AI whose safety is proved, not promised.
Long-term goal
AI systems that are safe and verifiable, with safety guarantees established by mathematical proof rather than anyone's assurance, so that no one's word needs to be trusted.
Current focus
The formal verification of the safety of neural networks and language models in everyday use.
A proof is only as good as the property it states, so part of the work is making safety properties precise enough to prove.
I’m an AI researcher and software engineer with three-plus years building production backend systems, plus peer-reviewed publications in computer vision, reinforcement learning, and generative models. I also build LLM systems on the side, most recently a natural-language-to-SQL system using RAG and vector search.
What I do
Research. My published work applied deep learning to medical imaging and agriculture (Frontiers, IEEE, Springer). Building and evaluating those systems is what made me want guarantees that go beyond test-set accuracy. My current focus is the formal verification of the safety of neural networks and language models.
Teaching. Lecturer, Dept. of Computer Science, American International University–Bangladesh (AIUB): programming, computer architecture, operating systems, software engineering, and web development. Lead developer of the department’s Web Development & Automation team, building systems to digitize and automate departmental activities, including the student course registration system (AIUB’s open-credit model) and the admission management system.
Software engineering. Three-plus years at Orion Informatics Ltd., building full-stack systems (Spring Boot, Angular, MS SQL Server, Apache Kafka) for Equiniti US’s shareholder/company management platform, including a Kafka messaging pipeline handling millions of shareholder votes per hour and a 60% reporting-query performance improvement.