Md Motaleb Hossen Manik

I develop methods for measuring AI reliability beyond task accuracy. In Unified Deployment-Aware Evaluation of Open Reasoning Language Models (TMLR, 2026), I evaluated seven model configurations across four benchmarks and three prompting strategies, covering 84 conditions and 19,992 examples, while jointly analyzing performance, uncertainty, latency, memory, prompt sensitivity, and deployment trade-offs.

I also developed HypothesisMed for structured reliability reporting, MILU for multimodal semantic evaluation, and earlier comparative work on ChatGPT versus DeepSeek for AI-based code generation.

I develop provenance and integrity mechanisms that make digital and AI-generated knowledge traceable, tamper-evident, reproducible, and securely exchanged. SlideChain provides blockchain-backed semantic provenance for structured VLM outputs over 1,117 lecture slides, with 100% tamper detection and deterministic reproducibility under controlled testing.

Earlier work includes my M.Sc. research on secure multi-party skyline queries and ShaEr, a privacy-preserving medical-data sharing framework.

I study how autonomous AI systems behave and act under permissions, policies, social interaction, and governance constraints. My workflow-agent research evaluates permission scope, semantic policy gating, and tamper-evident execution records; OpenClaw studies risky instruction propagation and norm enforcement across 14,490 autonomous agents; and ADAPT investigates adaptive governance in AI-mediated scholarly publishing.

The ADAPT work led to U.S. Provisional Patent Application No. 63/975,609.

I develop multimodal and interactive AI systems for education and medical knowledge workflows. I helped build MEDI-SLATE, comprising 1,117 slides and 262,182 narration tokens from a 23-lecture medical imaging course; contributed to ALIVE, a fully local interactive lecture system; and co-developed an optically emulated computed tomography scanner for hands-on medical imaging education.

My earlier work established a foundation in machine learning, NLP, and robustness, including Bangla and phonetic-Bangla sentiment analysis, unified sentiment and emotion recognition, adversarially robust text classification, and closed-domain question answering. Related methodological work includes N-ReLU, a stochastic extension of the ReLU activation function.