I build AI systems that can be trusted, and I test them by attacking them first. My PhD research covers LLM jailbreak defense, privacy attacks on VR users from encrypted traffic, and Riemannian-geometry defenses for EEG brain-computer interfaces. In 2026 at IBM Research Ireland, I built a Bayesian source-credibility layer for the FactReasoner fact-checking pipeline.
About Me
I am a Ph.D. candidate in Computer Science and Engineering at the University of Notre Dame, Indiana, USA. I build AI systems that can be trusted, and I test them by attacking them first. My dissertation focuses on learning trustworthiness from contextual metadata to build secure and privacy-preserving intelligent systems.
As a PhD Research Intern at IBM Research Europe, Dublin, I built ConTrust, a Bayesian source-credibility layer for IBM's proprietary FactReasoner pipeline. It beat six competing systems on balanced accuracy across three fact-checking datasets and exposed label bias in baselines that led on raw accuracy.
My research spans LLM jailbreak defense (GCG attack success cut to 12.5%), VR user privacy from encrypted network traffic (91% activity identification), and Riemannian geometry-based defenses for EEG brain-computer interfaces (cross-session re-identification cut from 90.8% to 15.6% and 9.9% on two public datasets, with no learned privacy model). You can find my publications on my Google Scholar profile.
Before joining Notre Dame, I worked as a Service Operations Engineer at bKash Limited, Bangladesh's largest fintech company, where I monitored mission-critical operations, applied software QA methodologies, and operated AWS infrastructure at scale.
Experience
Publications
Media Coverage
Research Projects
Automated defense pipeline against jailbreak and prompt-injection attacks without model retraining. Incorporates adversarial research with parallelized NLP-based summarization to dynamically adapt LLM responses.
Real-time recognition model classifying VR apps and in-app activities using encrypted network traffic. Wireshark-captured data processed with Random Forest, Decision Tree, and MLP architectures.
Ensemble-based framework modeling source reputation and trust in Retrieval-Augmented Generation systems. Uses NLP techniques and metadata analysis to quantify credibility and filter unreliable sources.
Training-free Riemannian transform that removes cross-session identity from motor-imagery EEG by equalizing each subject's mean covariance and tangent-space scatter. The worst of 5 attackers falls from 90.8% to 15.6% on BCI Competition IV-2a and to 9.9% on OpenBMI, at a cost of 7.0 and 2.1 points of decoding accuracy.
Sports Analytics
Open tools and case studies that turn match data into findings, with every number traced back to the data. Follow along on X: @SamitOTB.
Compares any football match with a team's other matches, at team and player level. Findings come from fixed, position-aware rules, and the optional AI-written summary is validated: every number in it must exist in the data, and claims the findings do not support are flagged. The live app needs a free Highlightly data key (steps are in its sidebar).
Barcelona won all 9 of their earlier La Liga games, then created 68% less xG against Real Madrid (1.19 against a 3.72 average). A player-level breakdown: Yamal shut out as a creator but not as a shooter, Raphinha's shots worth about a ninth as much, and a build-up that lost its two main passers.
Past Projects
During my B.Sc. at IUT, I conducted research on intelligent control systems and power engineering — applying optimization algorithms and neural networks to real-world energy systems problems.
Designed an AI-based online control scheme for regulating interconnected two-area thermal power systems using advanced metaheuristic and ANN-based approaches for real-time load frequency control.
Developed an enhanced gradient-based metaheuristic-optimized controller for load frequency control of a two-area AGC system. Benchmarked against conventional PID tuning methods.
Designed and implemented an automated vehicle access and exit control system for smart parking using a Finite State Machine architecture, fully described and simulated in VHDL.
B.Sc. thesis investigating PID and Fractional Order PID controller design for load frequency control of a two-area AGC system, tuned using optimization algorithms and artificial neural networks.
Conference Presentations
Academic Service
Voluntary Work & Community
Technical Skills
Education
Always open to research collaborations, academic conversations, and new opportunities.