Md Kamrujjaman Mobin
Building on-device AI at Samsung R&D Bangladesh while pursuing research in computer vision, embodied AI, and multimodal learning.
About Me
Research Interests
IELTS Academic: 7.0 Nov 2025
Education
Thesis: AI-Driven Health System for Mortality Prediction Using ECGs with Transformers and Selective State-Space Models. Focused on deep sequence modeling, representation learning, and AI-driven healthcare analytics.
Thesis PDFTechnical Skills
Professional Experience
Samsung R&D Institute Bangladesh
Software Engineer I · Dhaka, Bangladesh
On-device Android (Kotlin) agent that unifies data from multiple health platforms (Samsung Health, MyFitnessPal, Strava, etc.) via Google Health Connect and an MCP-style tool architecture, answering user queries with personalized insights and cross-app correlations. Fine-tuned Gemma 4 2B-IT for tool calling and compact system prompts, and deployed it with LiteRT-LM using on-device GPU acceleration to substantially reduce inference time.
Android app for longitudinal cognitive health monitoring via on-device ML. Multi-module Clean Architecture (Hilt DI, Room DB, WorkManager); passive pipelines for app usage, screen events, sleep, and call logs; on-device ONNX/TFLite inference over a 10+ biomarker feature extraction pipeline.
Clinical-grade SDK collecting 40+ sensor modalities from smartphone & Wear OS (PPG, ECG, EDA, BIA, accelerometer, sleep stages, app usage). KMP multi-module architecture with compile-time DI, SQLCipher-encrypted storage, GZIP sync, and AES file encryption for HIPAA-compliant data handling.
Wearable sensor calibration utility. Refactoring achieved 7.5% SAM score improvement, 20% CBO reduction, and 90%+ unit test coverage.
Research on non-invasive blood glucose estimation from PPG signals and multimodal Alzheimer's risk modeling targeting Samsung Health's global platform.
Training-Free Representation and Prompting for Event-Based Multimodal Understanding
Mobin, M.K., Shamil, M.S., & Arefeen, M.A.
Technical Report · EventBench General Understanding Track, Workshop on Event-Based & Multimodal Vision
Training-free system rendering event streams as polarity-coded, temporally ordered image grids for a frozen VLM (Qwen3.5-9B) with task-specific bin selection, cropping, and prompting. Ranked 9th of 27 teams (62.28% avg. across 5 tasks).
Research Projects
Active and ongoing research in embodied AI, neuromorphic computing, long-horizon video understanding, and multimodal & spatial reasoning.
Event-Based Vision & ANN–SNN Hybrid Perception
Investigating event-stream representations and temporal models for long-horizon egocentric video understanding on the newly released EventKitchen dataset (ECCV 2026). Earlier, surveyed 10+ papers (EventFly CVPR 2025, HsVT, SSM-based models) and ran experiments on N-EPIC Kitchens: voxel-grid generation, E2VID reconstruction, SSM+SNN temporal modeling. Proposing spike-native state-space models and noise-aware ANN-SNN hybrids.
GSA-VLN: General Scene Adaptation for Vision-Language Navigation
Replicated GR-DUET (ICLR 2025) on R2R/Matterport3D. Semantic-Aware Navigation with InfoNCE contrastive loss achieved +1.4% SR, +1.6% SPL. Experience Replay for continual learning: +0.7% late-episode SR.
MindCube Spatial Reasoning Challenge
Fine-tuned Qwen2.5-VL-3B-Instruct via QLoRA on MindCube using cognitive-map + chain-of-thought outputs. Built a 3-run weighted ensemble combining a custom SFT model and GRPO-trained RL model with self-consistency sampling. Structured outputs force explicit spatial reasoning before prediction; now extending this into a full research study.
Event-Based Multimodal Understanding Challenge
General Understanding Track at ECCV 2026 Workshop on Event-Based & Multimodal Vision. Developed event-stream MLLM solutions for spatiotemporal reasoning and scene understanding. Training-free approach with a frozen Qwen3.5-9B ranked 9th of 27 teams (62.28% avg. across 5 tasks).
Serialization Bottleneck in Tool-Use Agents
Measuring how text-serialized intermediate representations (polygons, graphs, tables, ASTs) degrade compositional LLM agent pipelines. Evaluating runtime-managed object handles with deterministic property summaries as mitigation.
Image Matching 2024 — Kaggle
Multi-model 3D reconstruction pipeline integrating COLMAP with deep feature matchers (LightGlue, SuperGlue, DKM, LoFTR).
ISIC 2024: Skin Cancer Detection — Kaggle
CNN-based multimodal models combining image and tabular data with advanced augmentation techniques for melanoma classification.
Projects
Personal and open-source projects across AI, full-stack, and embedded systems.
MCP Powered Weather AI
Conversational weather app with a custom Model Context Protocol (MCP) server and Gemini LLM. Supports standalone server and integrated client-server mode for terminal interactions.
AskProfessor — RAG Conversational AI
Lightweight PDF Q&A using open-source LLMs in a Retrieval-Augmented Generation framework. LangChain orchestration, ChromaDB vector store, SQLite chat history, scalable FastAPI backend.
LLM Fine-Tuning
Fine-tuned LLaMA 3.1, Mistral, InstructGPT, and DeepSeek AI via QLoRA with n-bit quantization and prompt engineering. Optimized with LoRA adapters and mixed-precision training.
Inventory Management System
Full-stack web app for real-time inventory, sales, and order tracking with responsive UI. Complete lifecycle — MySQL schema, Express.js server-side logic, EJS templating.
E-commerce Website with API
Modular e-commerce platform with RESTful APIs for product management and supplier delivery. Integrated real-time order processing and buyer/seller authentication.
Virtual Classroom Using OOP
Google Classroom-inspired simulation built on core Java OOP principles. Features user login, assignment uploads, and submission tracking via a MySQL backend.
IoT-Based Flood Detection System
IoT flood monitoring using Arduino, ultrasonic sensors, and ESP8266 for internet-based alerts. Real-time water level detection with automated email alerts on exceeding danger thresholds.
Get in Touch
Open to PhD & MSc opportunities, research collaborations, and interesting conversations.