Hi, I'm Assaduzzaman Munna — a Graduate Teaching Assistant at North South University, AI Researcher, and Nippon AI Dojo 2025 Fellow (Chowa Giken, Japan), with prior production experience at The Data Island. I conduct peer-reviewed deep learning research and build scalable, real-time edge inference systems.
I'm Assaduzzaman Munna — a CSE graduate from North South University with a rare combination: peer-reviewed AI research published at IEEE and under review in major international journals, paired with hands-on production engineering experience.
My work lives at the intersection of deep learning research and deployment reality. I don't just build models that score well on benchmarks — I build systems constrained by edge hardware, hospital imaging workflows, enterprise LLM architectures, and factory-floor latency requirements.
I was competitively selected for Nippon AI Dojo 2025 by Chowa Giken and AI Samurai Japan — working directly under Japanese AI industry mentors on real-world AI engineering problems. Currently, I serve as a Graduate Teaching Assistant in NSU's ECE Department while conducting peer-reviewed AI research, backed by prior production experience as an AI Engineer at The Data Island.
Whether I'm designing shared-private cross-modal attention mechanisms, compressing a ConvNeXt teacher model by 18.5× without accuracy loss, or developing agentic RAG architectures — I bring the same standard: precise, principled, and impactful.
From model architecture research to production deployment — a full-stack AI engineering skill set built through real research and industry practice.
From enterprise video analytics and agentic RAG architectures to embedded edge systems — built with production rigor.
Built and deployed highly optimized NVIDIA DeepStream pipelines for critical industrial safety applications, including low-latency fire and smoke detection. Combines spatial object detection with custom sliding-window Temporal Persistence Tracking (TPT) to eliminate false positives on live plant feeds.
Engineered an automated IP camera discovery and spatial mapping infrastructure to centralize large-scale industrial surveillance networks under The Data Island. Features multi-stream parsing, high-throughput GStreamer backends, and low-latency feature vector matching.
Built a low-latency Python/OpenCV backend capable of processing high-throughput inspection streams for industrial manufacturing lines. Performs real-time image alignment and defect segmentation with near-zero latency overhead.
Developed an intelligent agentic architecture utilizing Retrieval-Augmented Generation (RAG) and LLMs to automate semantic code search, dependency graph traversal, and multi-file contextual Q&A across complex software repositories.
Engineered a shared-private multimodal framework using context-aware cross-attention for joint emotion recognition and continuous sentiment prediction across 4 benchmark datasets. Resolves modality bias and improves cross-modal representations.
Engineered a knowledge distillation model reducing parameters by 18.5x for edge deployment, achieving 99.80% test accuracy with highly optimized inference speeds on an NVIDIA T4 GPU for automated renal calculi detection.
A research project evaluating reasoning capabilities in compact LLMs. Applies Parameter-Efficient Fine-Tuning (PEFT) using LoRA to adapt Qwen3-4B for advanced QA performance while analyzing reasoning pathways.
A local AI pilot serving SmolLM2-1.7B-Instruct on a Raspberry Pi 5. Utilizes multiprocessing to bridge real-time 60Hz Pymunk physics with LLM decision inference, executing evasive maneuvers, auto-hover locks, and safety overrides based on live UDP telemetry.
A compact hardware-software system designed using an STM32 microcontroller to monitor and display ECG signals, heart rate, and oxygen saturation levels. Integrates biosensing filters and displays real-time health data on a mobile dashboard.
From disciplined Git branch lifecycles and CI/CD quality gates to high-throughput real-time AI inference pipelines.
main = final/trusted
feature/* = dev work
PR = review + CI
CI/CD = test → build → deploy
Real metrics from real projects — no fabricated statistics.
Academic contributions advancing AI applications in medical imaging, affective computing, and multimodal alignment.
Submitted to Alexandria Engineering Journal — Currently under peer review
2026 IEEE 2nd International Conference on Secure IoT, Assured and Trusted Computing (SATC), Houston, TX, USA · DOI: 10.1109/SATC69565.2026.11542322
Submitted to IEEE SPICSCON 2026 — Currently under peer review
I'm currently open to full-time AI/ML engineering roles, research collaborations, and internships. I bring both research depth and practical engineering experience — a combination that's genuinely rare at this stage.