Current research, publications, intellectual property, funded work, appointments, teaching and academic service.
Topology-aware world model for electricity-grid reconfiguration that evaluates line-switching candidates through learned imagined rollouts and verifies promising actions using exact AC power-flow simulation — covering topology-changing actions, ranking of candidates, and identification of unsafe or poor candidates. Experiments are ongoing; no performance claims are made.
Investigating whether generation-time signals from autoregressive language models — particularly token-level logit and probability structure — can provide reliable evidence of output correctness, whether these signals distinguish reliable from incorrect or unsupported outputs, and how robust they are across models, tasks and decoding conditions.
Measuring harm amplification in multi-agent LLM systems — how harm introduced locally by a compromised or misaligned agent propagates and is amplified at the system level. A preprint is available on arXiv; the study is currently being expanded to additional tasks, with submission planned once that work is complete. Second lead on the project (with Rahman, M. H., Mahfuz, T., and Chakraborty, P.) — research ideation, experimental design and execution, and benchmark and evaluation methodology for robustness, controllability and adversarial failure modes.
ContributionProbes speech-recognition robustness via sparse, semantically targeted parameter bit-flips, characterizing controlled output changes and transferability.
Role: contributed to project ideation and methodology formulation, and to the project's ongoing technical discussions.
Role: contributed to early project ideation and built the malware-detection evaluation — how existing security tools respond to ML-specific attack behavior.
ContributionAnchor-calibrated trust estimation that recovers trusted and corrupted source cohorts when agreement-based methods become directionally ambiguous under corrupted majorities.
ContributionConsensus-based reward that negates malicious human-in-the-loop RLHF feedback by aggregating across a mixed-trust expert pool rather than rejecting updates outright — raising reward accuracy to ~73% from 48.78%, about 23% above prior work at DF = 10.
ContributionDifferentiable bit-level fault analysis for quantized vision–language models, studying controlled semantic steering under sparse parameter faults (BLIP, BLIP-2, GIT) — about 2.4× higher attack success than a prior baseline under a 20-bit budget.
SentriAlign: Training-Signal Integrity for LLMs
Commercialization of the COBRA consensus-reward line of work.
I-Corps Commercialization Activities
Customer discovery and exploration of practical applications and translation pathways for the COBRA research.
Private blinded multi-rater evaluation platform built for human validation of LLM-as-judge metrics in multi-agent AI-security experiments, with controlled rater assignment, randomized presentation, autosave/resume, structured annotation, and reproducible exports.
Remotely accessible heterogeneous Edge-AI testbed for AI, cybersecurity, edge-computing and resilience experiments — 21 NVIDIA Jetsons, 10 Raspberry Pis and 14 other single-board computers across two rack cabinets, with shared NAS storage, networking and rack/KVM infrastructure. Planned most of the architecture and led the deployment with another graduate researcher while guiding an undergraduate student.
Full-stack experimental tooling for structured review and downstream analysis of adversarial studies — repeatable AI-evaluation workflows via a Streamlit interface and Python export pipelines: structured scoring, diff-based review, and CSV export.
Resource-aware LLM execution across heterogeneous NVIDIA hardware, partitioning models layer-by-layer across Jetson Orin Nano, Xavier NX, AGX Orin and an RTX A6000, with models up to the 15B scale (StarCoder).