Loading
Iftekhar Ahmed
AI Engineer | Full Stack Dev
Hello 👋 I am Iftekhar, an AI engineer and Data Science master's student at the University of Helsinki. I like building AI products, tools, and software around ideas I find useful or interesting, and I enjoy experimenting to see what I can make out of them. On the research side, I mainly work in computer vision and am currently exploring Vision-Language Models (VLMs), Vision-Language-Action (VLA) systems, and Agentic AI.
Open to work
Internship, part-time, or full-time roles in —
AI EngineeringFull-StackResearch + ProductLLM ApplicationsAgentic Systems
Helsinki, Finland
Education:
M.Sc. in Data Science
University of Helsinki
Jul. 2026 – Present
Helsinki, Finland
Awarded 100% scholarship.
B.Sc. in Computer Science and Engineering
Leading University
Jan. 2020 – Jan. 2024
Sylhet, Bangladesh
CGPA 3.53 / 4.00 | Thesis on skin-cancer classification, published in IEEE Access (Q1).
Experience:
Investigator
Intex Research Lab · Part-time
Jun. 2023 – Present
Sylhet, Bangladesh
- Developed full-stack web apps end-to-end (frontend + backend), including Expert Tourism — a lab client project — and the lab website.
- Built and trained deep-learning models for medical imaging — published in IEEE Access (Q1) (skin-cancer classification) and Scientific Reports (Q1) (adversarial robustness of panoramic dental X-ray segmentation).
- Secured the NCA Cybersecurity Research & Innovation Pioneers Grant for a privacy-preserving federated-learning healthcare project.
- Promoted through Undergraduate and Graduate Research Assistant roles to Investigator.
Research Intern
Nov. 2025 – Sep. 2026
Remote
- Researched orthopantomogram (OPG) segmentation with the MIRA Wing — published Multi-Strategy Optimization of U-Net Variants for Orthopantomogram Segmentation at IEEE BECITHCON 2025.
Featured Projects:
Tics — Text · Image · Context · Search
ElectronTanstack RouterPythonCLIPFAISS
A local-first, privacy-focused desktop app for searching your personal image library using natural language, a reference image, or both combined. Everything runs on-device — no cloud, no telemetry.
Accountant AI
NextExpressPostgreSQLOpenAIAWSPython OCR
An AI-powered finance platform that automates bookkeeping, bank-statement conversion, expense tracking, and financial reporting. It turns plain text and uploaded documents into structured transactions, then surfaces analytics and accountant-grade reports for businesses, freelancers, and accountants.
Worked with:
AI/ML
PythonOpenCVPyTorchNumPyPandasTensorFlowHuggingFace
WebDev
ReactTypeScriptNextTailwindNodeExpressNestZodReactQueryTanstackAWS
Misc.
PostgreSQLMySQLMongoDBPrisma
Honors and Awards:
May, 2025
NCA Cybersecurity Research & Innovation Pioneers Grant
Awarded the prestigious NCA Cybersecurity Research & Innovation Pioneers Grant by the National Cybersecurity Authority of the Kingdom of Saudi Arabia for the research proposal "Privacy-Preserving Federated Learning Platform for the Healthcare Domain".
Oct, 2024
IBM TechXchange WatsonX Hackathon — Top 100
Achieved a top 100 ranking in the IBM TechXchange Watsonx Hackathon, earning a complimentary ticket to attend the IBM TechXchange Conference in Las Vegas, NV.
Oct, 2024
NASA Space Apps Challenge 2024 — Global Nominee
Our team ORBITUS was selected as one of 947 Global Nominees from 9,900+ teams worldwide in the NASA Space Apps Challenge 2024.
Featured Publications:
TFFM: Topology-Aware Feature Fusion Module via Latent Graph Reasoning for Retinal Vessel Segmentation
I. Ahmed, S. Absar, A. A. Sami, S. Sakib, D. Biswas, S. A. M. Mostafa
P2P-CV Workshop @ WACV 2026 |WACV
TL;DR:
Standard retinal vessel segmentation models produce fragmented, disconnected vessels, limiting reliable clinical analysis. Built a topology-aware model that preserves vascular connectivity using graph-based feature fusion and topology-driven loss functions—achieving SOTA performance while reducing vessel fragmentation by ~38%.
Multi-Model Attentional Fusion Ensemble for Accurate Skin Cancer Classification
I. Ahmed, B. Bushon Routh, M. S. Rahman Kohinoor, S. Sakib, M. Mahfuzur Rahman, F. Azzedin
IEEE Access 2024 |Q1
TL;DR:
Skin cancer classification models struggle with diverse lesion types, image artifacts, and dataset imbalance, limiting diagnostic reliability. Built an attention-enhanced ensemble combining ResNet50V2, MobileNetV2, and EfficientNetV2, along with robust preprocessing for artifact removal—achieving superior precision/recall and strong performance across challenging lesion classes.
News:
Featured Blogs:
February 5, 2026
Attention Mechanism — As I Understand It
I originally planned to write about the Transformer architecture directly, but since attention is the core innovation that powers Transformers, I decided to start with attention first. It took a while, but finally finished
August 5, 2025
Pixel-by-Pixel Image Creation: Understanding PixelCNN
This blog post aims to make PixelCNN easy to understand for beginners. We’ll start with an example to show the problem PixelCNN is designed to solve, then break down its core idea of using masked convolutions to model images pixel by pixel. From there, we’ll walk through its architecture step-by-step, including Type A and Type B masks, and build a simple PixelCNN from scratch using PyTorch. We’ll also discuss its strengths, limitations, and how it connects to similar ideas in other fields. By the end of this blog, you’ll have a clear understanding of what PixelCNN is, how it works, and how to implement it yourself.
July 25, 2025
The Game of Deception: Generative Adversarial Network (GANs)
The blog post explores Generative Adversarial Networks (GANs), which consist of two neural networks, the Generator and the Discriminator, competing in a dynamic process to create realistic data. Unlike Variational Autoencoders (VAEs), GANs produce sharp images by having the Generator create fake data while the Discriminator evaluates its authenticity. This adversarial interaction drives both networks to improve continuously, resulting in highly convincing outputs. The post explains the architecture and training of GANs using an analogy of a counterfeiter and a detective and discusses challenges like mode collapse and training instability. It concludes with a practical example of building a GAN using PyTorch and the Fashion-MNIST dataset, showcasing the generation of new fashion items.