Assaduzzaman Munna

Optimizing Intelligence. Scaling Architectures.

Bridging cutting-edge deep learning research with production-grade, edge-ready software. Specializing in computer vision models, knowledge distillation, multimodal temporal systems, and high-throughput deployment pipelines.

18.5x Model Compression
99.80% Renal CT Accuracy
3.75 CS CGPA (NSU)
⚡ latency: 2.4ms
Portrait photo of Assaduzzaman Munna, AI/ML Specialist
model:hybrid_cnn_vit
loss:0.0042
acc:99.80%

The Engineering Perspective

My technical core lies at the convergence of machine learning theory and system-level execution.

I am an AI/ML Engineer driven by the challenge of deployment efficiency. I believe artificial intelligence shouldn't just run on high-end remote server racks—it must be compressed, optimized, and ready to deploy at the edge to impact real-world latency-constrained operations.

Currently, I serve as a Junior AI Engineer at The Data Island, where I construct data processing pipelines and integrate robust machine learning frameworks for enterprise systems.

Through academic endeavors at North South University (CSE Department, CGPA: 3.75) and selective industry programs like the Nippon AI Dojo 2025 (mentored by Japanese industry leads from Chowa Giken and AI Samurai Japan), I have cultivated deep expertise in model distillation, neural network compression, and low-latency computer vision backend services.

import torch
import torch.nn as nn
class StudModel(nn.Module):
def __init__(self):
super().__init__()
self.vit = MobileViT()
def forward(self, x):
return self.vit(x)
# Distilling knowledge...
loss = KDDiv(alpha=0.9)
opt = AdamW(model.parameters())
epoch_acc = 0.9980
parameters_reduction = 18.5x
status: deploying_to_edge
jetson_inference: running...
latency_average = 2.4ms
import torch
import torch.nn as nn
class StudModel(nn.Module):
def __init__(self):
super().__init__()
self.vit = MobileViT()
def forward(self, x):
return self.vit(x)
# Distilling knowledge...
loss = KDDiv(alpha=0.9)
opt = AdamW(model.parameters())
epoch_acc = 0.9980
parameters_reduction = 18.5x
status: deploying_to_edge
jetson_inference: running...
latency_average = 2.4ms
Assaduzzaman Munna at work

Technical Matrix

Quantified competencies categorized by specialization layers.

Deep Learning & AI

PyTorch92%
TensorFlow & Keras80%
Hugging Face Transformers85%
Computer Vision (OpenCV)88%

Edge & Systems

NVIDIA DeepStream78%
Docker Containers85%
MLOps & Pipelines75%
Linux & Bash Shell82%

Engineering Languages

Python95%
C++ Programming80%
SQL Database Queries85%
Git & Version Control90%

Full-Stack Integration

ReactJS / NextJS82%
NodeJS Backend75%
HTML5 / CSS3 Layouts90%
RESTful APIs Integration86%

Timeline of Impact

Industry roles and specialized training programs focused on high-performance AI deployment.

Junior AI Engineer

April 2026 - Present
The Data Island
  • Design, develop, and integrate enterprise AI frameworks, handling massive scale data ingestion.
  • Establish high-efficiency, reliable data processing pipelines for machine learning operations.
  • Collaborate closely with core developers to optimize deep learning training routines, significantly reducing inference compute times.

On-the-Job AI Trainee

Sep 2025 - Jan 2026
Nippon AI Dojo 2025 (Chowa Giken & AI Samurai Japan)
  • Selected as one of the top candidates for a highly intensive AI engineering and model optimization program.
  • Engineered production-level vision systems and model pipelines under direct training and review of industry architects in Japan.
  • Successfully completed hands-on challenges regarding neural compression, quantization, and real-time edge processing constraints.

Undergraduate Teaching Assistant

June 2025 - Present
North South University (ECE Dept.)
  • Conduct review sessions and facilitate laboratory exercises in core computer science and engineering coursework.
  • Provide direct instruction and programming mentorship to over 80+ undergraduate engineering students.
  • Support faculty in structural curriculum reviews and evaluate engineering course deliverables.

Research & Publications

Peer-reviewed publications and manuscripts demonstrating optimized diagnostic systems and deep architectures.

Hybrid CNN-MobileViT Model with Knowledge Distillation

Submitted to SPICSCON 2026 (Under Review)

Engineered a compact hybrid student model combining spatial CNN features and transformer-based self-attention loops for efficient Renal Calculi Detection in CT Images. Used knowledge distillation to transfer features from a massive ConvNeXt-Tiny teacher.

Compression: 18.5x parameters reduction Inference Accuracy: 99.80% Compute: 9 mins training on NVIDIA T4

Bone Fracture Detection Using Vision Transformers

2026 IEEE 2nd International Conference on Secure IoT (SATC)

Conducted a comparative analysis of Pooling-based Vision Transformer (PiT) and CaFormer architectures against standard convolution models using a dataset of 4,083 annotated clinical X-ray images. Demonstrated superior generalization vectors and robustness in localized anomaly classification.

Validation Accuracy: 97.51% Dataset: 4,083 annotated X-rays Framework: PyTorch & Hugging Face

Featured Systems

Industrial and experimental software systems engineered for efficiency, concurrency, and real-world deployment.

MIST-ER

Deep Learning Architect

A multimodal temporal emotion recognition architecture processing high-frequency audio, video, and text streams concurrently to perform micro-expression classification.

  • Achieved 54% categorical accuracy on the highly challenging and noisy MESC dataset.
  • Designed a bidirectional attention synchronization layer matching video and audio frames in time.
  • Implemented parameter quantization, allowing real-time deployment on edge devices.
PyTorch Multimodal Fusion Attention Layers Edge Optimization

Foil Stamping Inspection

System Architect

A production-ready industrial automation computer vision system designed to inspect foil stamping lines for high-throughput manufacturing facilities.

  • Designed high-performance backend pipelines using Python and multithreaded OpenCV frameworks.
  • Optimized algorithms to handle sub-millisecond inspection cycles with extreme constraint boundaries.
  • Deployed on client site manufacturing lines, improving structural inspection defect rates.
Python OpenCV Multithreading Industrial Automation

Establish Contact

I am open to discuss engineering roles, research collaborations, or technical architecture challenges. Reach out via direct email or professional networks.

Please enter a valid identifier.
Please enter a valid routing email address.
Please input a message payload longer than 10 characters.