Chashi Mahiul Islam

Ph.D. Candidate in Computer Science · Florida State University

Ph.D. candidate in AI/ML building reliable, robust, and well-evaluated LLM, agent, and multimodal systems.

Seeking full-time Machine Learning Engineer, Applied Scientist, or Research Scientist roles in LLMs, agents, and vision-language models (start 2027).

+1 (850) 559-8366  ·  chashi.buet@gmail.com  ·  linkedin.com/in/chashi-bd  ·  github.com/meetchashi  ·  meetchashi.github.io

Education
Florida State University Tallahassee, FL
Ph.D. Candidate, Computer Science (Advisor: Dr. Xiuwen Liu)
Research interest: Robustness of Large Multimodal Models
December 2026 (Expected)
M.Sc. in Computer Science, CGPA 3.97/4.00 · August 2025
Bangladesh University of Engineering and Technology Dhaka, Bangladesh
B.Sc. in Computer Science and Engineering
April 2019
Work Experience
Chewy Jun 2026 – Aug 2026
Machine Learning Engineer Intern II, Bellevue, WA
  • Built a context-aware LLM personalization framework for Chewy's shopping assistant (context injection, proxy orchestration, prompt optimization); raised personalization scores 20% (3.5 → 4.2) on 4 rubrics.
  • Used rubric-based evaluation and failure-mode analysis to guide iterative prompt optimization.
Meta May 2025 – Aug 2025
Software Engineer Intern, Seattle, WA
  • Integrated vision features into a prompt generation model for large-scale feed ranking, improving click-rate prediction AUROC by 3.4%; built feature selection and scalable data pipelines to ship new features.
Florida State University Jan 2021 – Present
Graduate Research and Teaching Assistant · Dept. of Computer Science, Tallahassee, FL
  • Conducting research on robustness and explainability of multimodal AI systems, focusing on vision-language models and LLMs and applying machine learning techniques to enhance AI reliability and security.
  • Led recitations, developed assessments, and supported 60–90 students in Secure, Parallel & Distributed Programming, Data Structures, Advanced Python, and OOP with C++.
CoKreates Limited Aug 2019 – Dec 2020
Software Developer, Dhaka, Bangladesh
  • Built front-end and REST APIs for an ERP admin panel, Human Resource (HR) module, and centralized logging (Angular, Spring Boot).
Selected Publications
  • NeurIPS 2026 (Accepted): Neural Expansion: A Unified Mechanism for How Deep Neural Networks Generalize. Islam, C.M., Jacob Chacko, S., Nishino, M., Zhang, C., Liu, X. Geometric account of generalization via expansion of prediction-stable regions. (paper, code)
  • WCCI-IJCNN 2026 (Oral): Numerical Instability and Chaos: Quantifying the Unpredictability of Large Language Models. Islam, C.M., Villarreal, A., Nishino, M., Salman, S., Liu, X. Regimes of stability and rounding-induced chaos. (paper, code)
  • ISVC 2025 (Oral): NeuroShield-ViT: Mechanistic Understandings of Representation Vulnerabilities and Engineering Robust Vision Transformers. Islam, C.M., Jacob Chacko, S., Nishino, M., Liu, X. PGD-100 accuracy 15.6% → 72.3%, no retraining. (paper)
  • IROS 2024 (Oral): Malicious Path Manipulations via Exploitation of Representation Vulnerabilities of Vision-Language Navigation Systems. Islam, C.M., Salman, S., Shams, M., Liu, X., Kumar, P. Adversarial navigation redirection, noise-sensitivity detection. (paper, code)
Research Experience
  • Numerical Instability of Large Language Models (WCCI-IJCNN 2026, Oral): Discovered that microscopic input perturbations cause up to 10-million-fold amplification in LLMs, collapsing logit margins and triggering unpredictable token flips. Quantified regimes of stability and rounding-induced chaos. (paper, code)
  • Robust Vision Transformers: Understanding and Mitigating Representation Vulnerabilities (ISVC 2025): Examined how small input changes affect layer-wise representations in Vision Transformers. Proposed NeuroShield-ViT to neutralize vulnerable neurons and enhance adversarial robustness without retraining, raising PGD-100 accuracy from 15.6% to 72.3%. Classification accuracy also improved by 71.6% against IGO attacks on ImageNet1K. (paper)
  • Adversarial Exploitation in Robot Vision-Language Navigation (IROS 2024): Developed a vision-language navigation (VLN) framework for embodied robotic agents, analyzing perception-to-action pipelines within vision-language-action (VLA) architectures. Studied how representation shifts propagate through navigation policy networks, affecting task planning and real-world embodied decision making. Achieved 91.7% landmark-conditioned path redirection and proposed a noise-sensitivity-based detection mechanism improving navigation policy stability by 96.3%. (paper, code)
  • ActivationGuard: Prompt Injection Detection in LLM Agents (under review): Lightweight detector on frozen LLM hidden states with logistic regression; expanded training coverage raised attack recall from 17.2% to 93.8% under 5% FPR validation budget - at 0.979 AUROC.
  • Adversarial Attacks on LLMs via Regularized Relaxation (Information Sciences, 2026): Co-developed regularized continuous suffix optimization with token mapping; higher attack success on five LLMs, ≈100× faster than GCG. (paper, code)
  • Instruction Strategies for Tool-Grounded AI Agents (MCP, agent evaluation; 2026): 180 cybersecurity-agent runs across four guidance levels; no significant success-rate gain as context grew from 591 to 36,001 tokens. (paper)
  • SpatialViLT (ISVC 2025): Fused depth maps, 3D coordinates, and edge representations into a vision-language model via multi-task learning to improve spatial grounding and cross-modal alignment in embodied intelligence systems. Enhanced 3D scene understanding for downstream robotic navigation and manipulation, improving visual spatial reasoning benchmark performance by 2.1%. (paper)
  • Hallucination in Multimodal LLMs (ISVC 2025): Traced how vision-encoder embeddings propagate hallucination vulnerabilities and identified representation-level failure modes behind systematic visual misperception (e.g., DeepSeek). (paper)
Selected Projects
  • De Novo Peptide Sequencing Reliability (Transformers, Pyteomics; 2025): Stress-tested Casanovo and π-PrimeNovo with physics-preserving spectral perturbations, exposing 73.6% and 95.8% prediction instability and revealing fundamental reliability gaps in de novo sequencing models. (code)
  • Private LLMs for Unstructured Text (Transformers, PEFT/LoRA, BitsAndBytes; 2023): Hosted LLMs locally to protect data privacy; fine-tuned summarizers and classifiers with LoRA and prompt engineering, avoiding full-model retraining. (code)
  • Low-Resolution Deepfake Detection (Xception, super-resolution; 2022): Evaluated state-of-the-art detectors on low-resolution video and added super-resolution, improving test accuracy by 19% on FaceShifter. (code)
  • Multivariate Time Series Classification (PyTorch, scikit-learn, XGBoost, tsfresh; 2023): Built a predictive XGBoost model to forecast intracerebral hemorrhage using time-series data and PubMed radiology reports. Achieved 74.11% accuracy and 75.51% precision, aiding hospital decisions by identifying high-risk patients. (code)
  • Comparison of Summarization Techniques Using Language Models (Transformers, NLTK; 2023): Investigated transformer models for medical report summarization, comparing BERT, GPT-2, T5, and BART. Addressed input size limits and domain-specific vocabulary to improve summarization effectiveness. (code)
  • Image Style Transfer with VGG Network (PyTorch; 2021): Implemented image style transfer using convolutional neural networks. Used a pretrained VGG-19 network to extract content and style representations and generate a new target image. (code)
Technical Skills
LLMs and Agents LoRA/PEFT, quantization, prompt engineering, RAG, LLM evals, prompt-injection defense, multimodal LLMs, hallucination analysis, LangChain, LangGraph, AutoGen, CrewAI, MCP
ML PyTorch, Hugging Face Transformers, computer vision, ranking systems, recommender systems, feature engineering, A/B testing, model evaluation, representation learning, scikit-learn, XGBoost, adversarial robustness, interpretability
Engineering Python, C/C++, SQL, NumPy, Pandas, data pipelines, Weights & Biases, AWS, Docker, Git, Linux, PostgreSQL
Relevant Courses
  • Graduate: Introduction to Data Science, Data Mining, Advanced Database, Advanced Algorithms, Analytic Methods, Deep Learning and Reinforcement Learning Fundamentals, Data and Computer Communication.
  • Undergraduate: Python, C, C++, OOP, Digital Signal Processing, Computer Graphics, Software Development, Discrete Mathematics, Probability and Statistics, Database Management System, Image Processing and Pattern Recognition.
Leadership and Awards
  • Dean's Award for Doctoral Excellence, FSU College of Arts and Sciences (2025, 2026)
  • Naaman Franklin Faile Jr. Graduate Fellowship (2025, 2026)
  • GSRC and COGS Travel Grants, Florida State University, for ISVC 2025 in Las Vegas. Supported by the Graduate Student Resource Center (GSRC) and the Congress of Graduate Students (COGS).
  • COGS Travel Grant, Congress of Graduate Students, Florida State University, for IROS 2024 in Abu Dhabi.
  • ACM at FSU Fall Programming Contest: 1st Runner-up (2024), 2nd Runner-up (2023)
  • Cultural Secretary, Bangladesh Student Association – FSU (2021)
  • President, BUET Film Society (2018)