Md Mofazzal Hosen
Hi, I'm Mofazzal, a Computer Science and Engineering graduate from Khulna University of Engineering & Technology (KUET), Bangladesh. My research interests are medical image analysis, small-object detection, model calibration, and reproducible evaluation.
My work explores how machine learning can support medical image analysis, with an emphasis on reliable predictions and rigorous evaluation. My undergraduate thesis, supervised by Prof. Dr. Sk. Md. Masudul Ahsan, investigated tiny-human detection in aerial search-and-rescue imagery. I also co-authored a study of exercise heart-rate estimation from wrist photoplethysmography signals.
At MAEKNIT Inc., I develop Python-based agentic systems that turn user requests into controlled tool calls and task execution within Odoo workflows. Previously, at AlgoSoft Technologies, I worked on enterprise export/import software with AI-powered insights and a multi-tenant architecture for distributors and resellers across Europe. I am interested in graduate research opportunities that connect careful evaluation with useful learning systems.
Medical image analysis · Small-object detection · Model calibration · Reproducible evaluation
News
Research
My work focuses on visual recognition and physiological signals, with an emphasis on evaluation under limited data and distribution shift.
Manuscripts under review
Exercise Heart Rate Estimation Using Range-Covering Deep Learning on Wrist Photoplethysmography Signals
Submitted to ICCIT 2026 · 2026
Submitted · under review
Range-Covering Deep Learning addresses scarce high-heart-rate training coverage by adapting a 76K-parameter wrist-PPG estimator with distribution and spectral targets. Adding 59 minutes of treadmill recordings reduced held-out MAE from 18.0 to 11.9 bpm and reduced error by 19%–40% across five external datasets, while held-out PPG-DaLiA subject error did not change detectably.
Lightweight Tiny-Aware Multi-Scale Attention for Aerial Human Detection in Search-and-Rescue Imagery
Submitted to ICCIT 2026 · 2026
Submitted · under review
CBAM attention, a stride-4 P2 branch, and tiny-aware training assignment for YOLO11m. Official C2A AP50 increased from 0.8432 to 0.8533. Across three seeds, scene-disjoint recall for sub-8-pixel targets improved by 1.62 ± 0.49 percentage points, with lower recall for larger targets.
Ongoing research
Two-View Skin-Lesion Classification with Multi-Expert Aggregation
ISIC MILK10k 2025 challenge · KUET research collaboration · 2026
Ongoing research · live leaderboard result, not a final award
For paired clinical and dermoscopic images, team perseverance combines ConvNeXt-B representations, SADE multi-expert aggregation, and a rate-transferred decision layer. On October 20, 2026, the live leaderboard showed rank 13, mean Dice 0.569, and macro AUC 0.881. Ongoing work tests image fusion and class imbalance with lesion-level cross-validation, class-wise metrics, and bootstrap comparisons.
Experience
Research Member, Skin-Lesion Classification
KUET, Department of Computer Science & Engineering · Khulna, Bangladesh
Jointly developing two-view ConvNeXt-B models with multi-expert aggregation for paired clinical and dermoscopic images.
- Co-developed ViT-GPT2 and BLIP baselines in an earlier vision-language pipeline.
- Shared work includes lesion-level splitting, class-wise evaluation, calibration, and risk-coverage analysis.
Junior Software Engineer
MAEKNIT Inc. · New York, USA · Remote
Developing Python-based agentic systems that turn user requests into controlled tool calls and task execution within Odoo ERP workflows.
- Extend production ERP workflows with maintainable Odoo modules and server-side business rules.
- Implemented LLM tool calling and task execution through a dictionary-based registry that maps controlled tool requests to application functions.
Full-Stack Software Developer
Algosoft Technologies Limited · Bangladesh · Remote
Developed enterprise export/import software with AI-powered insights and a multi-tenant architecture for distributors and resellers across Europe.
- Shipped 37 Next.js/React routes across 8 modules and integrated more than 50 REST endpoints with caching and error boundaries.
- Built an OpenAI-assisted review tool with confidence flagging that reduced manual report drafting by approximately 70%.
- Delivered a production RAG chatbot across 8 knowledge categories and Python pipelines for PDF/web data extraction.
Software Developer (Intern)
AlgoSoft Technologies Limited · Bangladesh · Remote
Contributed to the ERP/DMS platform before continuing in a full-stack development role.
Education
B.Sc. in Computer Science & Engineering
Khulna University of Engineering & Technology (KUET)
Thesis: MSA-YOLO, tiny-human detection in aerial search-and-rescue imagery. Supervisor: Prof. Dr. Sk. Md. Masudul Ahsan.
Honors
KUET Technical Scholarship
Merit scholarship awarded to the top 50% of the class in four academic sessions.
Selected projects
Sentinel
2026An autonomous incident-response system where AI investigates and proposes, while deterministic Java guardrails retain authority.
Celloscope
2026A FastAPI service for Bengali/English transcription and structured lab-report extraction with deterministic offline replay.
PharmaPlus
2024–2026A two-sided Android pharmacy app with customer checkout, a live admin catalog, design patterns, automated tests, and CI builds.
Technical skills
Languages
Product engineering
Data & systems
Applied AI
Research evaluation
Teaching & service
Technical Lead
K-MINDS, Machine Intelligence & Data Science Community
Organized classes and weekly workshops, taught computer vision, managed the website team, and coordinated work across teams.
General Secretary
HACK, Hardware Acceleration Club, KUET
Organized regular classes and weekly workshops for the Hardware Acceleration Club.
Organizing Committee
BITFEST 2025, KUET CSE National Fest
Sponsorships, social media, outreach and administration. Helped run the hackathon and LFR event and contributed to hackathon problem setting.
Contact
For research opportunities and collaboration, please reach me at mofazzalhossen874@gmail.com.