Arafat Hossain Ankon AI Engineer & Researcher
Portrait of Arafat Hossain Ankon
Dhaka, Bangladesh Software Engineer

Arafat Hossain Ankon

AI Engineer · Medical AI Researcher · Founder, Digital Garage

I develop machine learning systems and research trustworthy evaluation methods for clinical AI. My work focuses on shortcut learning, domain generalization, and data leakage in medical image classifiers.

Download CV Explore Research
LinkedIn GitHub Email
Undergraduate GPA
3.92 / 4.00

Dean’s Award recipient (3 consecutive semesters)

Degree & Major
B.Sc. in Software Engineering

Specialization in Data Science · DIU Class of 2025

Core Research Focus
Medical Vision

Shortcut learning & cross-domain validation

Ventures & Leadership
Founder

Digital Garage · AI & Software Solutions

Credentials & Honors

Academic foundation & recognition

Formal software engineering education with intensive specialization in data science, predictive modeling, and algorithm design.

Daffodil International University

B.Sc. in Software Engineering

Major in Data Science · Class of 2021 — 2025

3.92 / 4.00
Undergraduate CGPA
  • Core Coursework: Machine Learning, Deep Learning, Computer Vision, Natural Language Processing, Distributed Systems, Data Structures & Algorithms, Database Engineering, Linear Algebra & Statistics.
  • Capstone & Thesis: Cross-domain validation methodology and shortcut mitigation in medical imaging classifiers.

Awards & Distinctions

Dean’s Award for Academic Excellence

Achieved a perfect 4.00 / 4.00 CGPA across three consecutive academic semesters in Software Engineering.

3-Minute Thesis (3MT) Competition Finalist

Selected as a university finalist for communicating complex medical AI research hypotheses clearly to a multidisciplinary audience.

Academic Research

Medical imaging, shortcut learning
& trustworthy AI

Investigating critical failure modes in deep learning evaluation: determining whether reported clinical metrics reflect true diagnostic generalization or hidden data leakage and spuriously correlated features. Investigating shortcut learning, cross-domain validation, and data leakage in clinical AI.

Research Manuscript Primary Investigator

Standard AUC Evaluation Cannot Detect Shortcut Learning in Medical Image Classifiers: A Controlled Cross-Domain Study

Investigates why standard within-distribution Area Under the ROC Curve (AUC) fails to detect when convolutional neural networks latch onto spurious background artifacts (shortcuts) rather than true clinical pathology. We demonstrate cross-domain performance collapse in external cohorts despite near-perfect in-distribution AUC. Investigates why standard in-distribution AUC fails to detect shortcut artifacts in CNNs, resulting in cross-domain diagnostic collapse.

Shortcut Learning Cross-Domain Generalization AUC Evaluation Limits Diagnostic Reliability
Methodology & Clinical Implications
Problem Statement

In medical AI diagnostics, deep networks frequently leverage non-pathological cues (e.g. scanner borders, hospital tokens, patient orientation tags) as shortcuts. While internal test sets reward these artifacts with high AUC scores, clinical deployment across distinct hospitals results in severe diagnostic failures.

Controlled Experimental Setup

We conducted a controlled cross-domain stress test isolating pathological lesions across multiple distinct imaging distributions. Models trained with standard ERM (empirical risk minimization) showed up to 28% drop in target domain classification while maintaining >0.96 internal validation AUC.

Implications for Clinical AI

Advocates for multi-domain out-of-distribution benchmark protocols and saliency-guided feature ablation before clinical deployment readiness is declared.

Research Manuscript Methodology Study

The Illusion of Accuracy: Quantifying Augmentation-Induced Data Leakage in Medical Image Classification

Quantifies how pre-split data augmentation and subtle patient overlap inadvertently inject shared spatial priors across train and test partitions. This work establishes rigorous partitioning guidelines to prevent optimistic performance overestimation in medical vision benchmarks. Quantifies how pre-split data augmentation injects shared spatial priors across partitions, establishing rigorous clinical split guidelines.

Data Leakage Quantification Pre-Split Augmentation Biases Patient-Level Stratification Validation Protocols
Methodology & Solution Guidelines
Key Findings

When augmentation operations (e.g. affine transformations, elastic distortions, cutouts) are applied prior to rigorous patient-level cross-validation splits, latent feature leakage inflates reported F1 and AUC by 12% to 19%.

Methodological Solution

We present an automated dataset audit pipeline that detects near-duplicate perceptual feature embeddings across splits and enforces strict patient-boundary isolation.

Selected Projects

Applied machine learning
& production systems

Translating cutting-edge computer vision, voice AI, and multi-modal models into robust real-world applications.

Computer Vision · Medical AI

OralScan

  • Early oral cancer detection pipeline using CNNs and Vision Transformers
  • Dual-backbone architecture evaluated with rigorous 5-fold cross-validation
  • Preprocessing with contrast enhancement and balanced data augmentation
PyTorch Vision Transformers CNN Scikit-Learn
View Source Code & Experiments
Conversational AI · Real-Time Voice

TableTalk AI

  • Low-latency voice ordering agent with sub-second auditory response
  • Bidirectional audio streaming using LiveKit WebRTC and Gemini voice pipeline
  • RAG retrieval for dynamic menu lookups, allergens, and order cart extraction
Python LiveKit Gemini RAG NLP
View Source Code & Pipeline
AI Automation · Content Pipeline

WordPress Blog Automation using n8n

  • Developed an end-to-end AI content automation pipeline
  • Integrated Perplexity AI, Gemini, and WordPress REST API
  • Automated content validation and publishing workflows
n8n Perplexity AI Gemini WordPress REST API Webhooks
GitHub Repository & Workflow
Venture & Product Studio

Digital Garage

Architecting digital solutions and AI-augmented workflows to modernize business operations and deploy practical machine learning software.

Explore Digital Garage
Career Path

Professional experience & teaching

Combining startup leadership, hands-on production machine learning engineering, and technical education.

2024 — Present Active

Founder & Technical Lead

Digital Garage

Overseeing end-to-end product strategy, technical architecture, and execution. Designing software systems and tailored workflow automations that solve concrete business challenges.

  • Architected scalable web applications and business intelligence integrations for clients
  • Engineered machine learning pipelines for automated document processing and decision support
Aug 2026 — Present Active

Teaching Assistant

Ostad

Mentoring cohort students in AI, machine learning, and automation. Guiding learners through foundational mathematics, feature engineering, neural network debugging, and real-world project development.

  • Facilitating hands-on live code sessions covering PyTorch, Scikit-learn, and computer vision
  • Conducting one-on-one code reviews and debugging sessions to reinforce ML engineering best practices
Jun 2025 — Present Active

Junior AI Engineer

Join Venture AI

Developing production machine learning models and intelligent data systems. Working collaboratively with cross-functional engineering teams to implement NLP and RAG pipelines.

  • Built automated exploratory data analysis (EDA) and data validation pipelines for tabular and text corpora
  • Integrated retrieval-augmented generation (RAG) workflows into enterprise production environments
Technical Arsenal

Skills & engineering technologies

Tools, frameworks, and methodologies applied across academic research pipelines and production deployments.

Machine Learning & Computer Vision

Model development, deep learning architectures, and rigorous validation.

PyTorch TensorFlow CNNs Vision Transformers Transfer Learning Cross-Validation Feature Engineering Scikit-Learn Medical Imaging

Language Models & Conversational AI

Natural language processing, audio pipelines, and generative systems.

NLP RAG Architectures LangChain LangGraph Vector Databases Prompt Engineering LiveKit WebRTC Gemini API

Programming & Scientific Data

Languages and scientific computing libraries for analysis and modeling.

Python SQL JavaScript / TypeScript Pandas NumPy Power BI Matplotlib Seaborn Jupyter

Databases & Infrastructure

Production hosting, data persistence, and CI/CD pipelines.

PostgreSQL Supabase Docker Git & GitHub GitHub Actions Linux Bash REST APIs
Get in Touch

Let's build something impactful together.

I am open to academic research partnerships, medical AI collaborations, technical speaking invitations, and select machine learning engineering consulting.

Direct Email Inquiry Connect on LinkedIn Explore GitHub Repositories Download Complete Curriculum Vitae