Md Ashik Khan

Computer Vision · Medical Imaging · Video Action Recognition

Senior Software Engineer, Auptimate  ·  M.Tech CSE, IIT Kharagpur

Portrait of Md Ashik Khan

About

I am a Senior Software Engineer at Auptimate, where I develop AI-driven systems for document processing and e-signature platforms. I completed my M.Tech. in Computer Science and Engineering from IIT Kharagpur in 2023, where I worked as a Research Assistant at the Computer Vision and Intelligence Research (CVIR) Lab under Dr. Abir Das.

My research is in computer vision and machine learning. I am interested in semi-supervised learning, resource-efficient methods, and trustworthy AI, such as detecting data leakage and mismatched data, especially in video action recognition and medical imaging. For my M.Tech thesis, I benchmarked six video models, including 2D CNNs, 3D CNNs and video transformers, on 14 datasets from six domains, and ran more than 400 experiments comparing Kinetics-400 and Moments in Time pretraining.

My work is connected by two ideas: getting strong results with small models, few labels and limited compute, and carefully checking whether those results are real. My first-author paper at MLHC 2026, SHIFT-M3, checks whether an LLM-generated ECG interpretation and a clinical report describe the same patient before fusion. In WICA-Net-M, we found that 857 of the 1,000 test images in the BRISC benchmark also appear in the main public brain tumor dataset. In my chest X-ray work, small adapters on a frozen encoder beat full fine-tuning (AUROC 0.908 vs 0.770) with about 40× fewer trainable parameters.

Research Interests

Computer Vision Medical Imaging Video Action Recognition Multimodal Learning Efficient ML Trustworthy AI

News

  • Paper "WICA-Net-M: MRI-Based Brain Tumour Classification Using a Lightweight Wavelet-Integrated Coordinate Attention Network with Frequency-Aware Learning" accepted in Computers (MDPI).
  • Paper "Human Activity Recognition Under Low-Light Conditions Using Frozen CLIP with Lightweight Adaptation" accepted in Discover Artificial Intelligence (Springer Nature).
  • Paper "SHIFT-M3: Pre-fusion Alignment-based Consistency Screening for Multimodal ECG Record Integrity" accepted at MLHC 2026.
  • Served as a reviewer for IEEE Access.
  • Served as a reviewer for IEEE Access.
  • Promoted to Senior Software Engineer at Auptimate.
  • Paper "Fixed-Threshold Evaluation Reveals the Forensic-Semantic Spectrum in AI-Generated Image Detection" accepted at IEEE 28th ICCIT 2025.
  • Paper "Fixed-Budget Parameter-Efficient Training with Frozen Encoders Improves Multimodal Chest X-Ray Classification" accepted at IEEE 28th ICCIT 2025.
  • Served as a reviewer for International Conference on Intelligent Data Analysis and Applications (IDAA 2025).
  • Paper "Comparative Analysis of Resource-Efficient CNN Architectures for Brain Tumor Classification" published at ICCIT 2024.
  • Joined Auptimate as Software Engineer.
  • Completed Master of Technology in Computer Science and Engineering, IIT Kharagpur.
  • Started as Research Assistant at CVIR Lab, IIT Kharagpur.
  • Awarded ICCR Scholarship under the Bangladesh Scholarship Scheme for M.Tech at IIT Kharagpur.

Publications

SHIFT-M3 dual-stream architecture diagram

SHIFT-M3: Pre-fusion Alignment-based Consistency Screening for Multimodal ECG Record Integrity

Md Ashik Khan, Md Nahid Siddique

Machine Learning for Healthcare (MLHC) 2026, Proceedings of Machine Learning Research (PMLR), vol. 340, pp. 815–834

Multimodal clinical AI assumes the waveform, report and metadata attached to a record belong to the same patient. SHIFT-M3 screens records before fusion by comparing an LLM-generated ECG interpretation with the clinical summary. On 784,680 records it detects 97.6% of complete swaps and 90.3% of partial swaps.

WICA-Net-M architecture diagram

WICA-Net-M: MRI-Based Brain Tumour Classification Using a Lightweight Wavelet-Integrated Coordinate Attention Network with Frequency-Aware Learning

Md Ashik Khan, Abu Saleh Musa Miah, Md Abdur Rahim, Jungpil Shin, Mohd Nizam Husen

Computers (MDPI), vol. 15, art. 586, 2026 · doi: 10.3390/computers15090586

A 2.47M-parameter CNN trained from scratch, with a gated Haar wavelet transform and Coordinate Attention, evaluated under a leakage-aware, three-seed protocol. It reaches 95.25% accuracy on the deduplicated split, matching five ImageNet-pretrained baselines, and an overlap audit shows 857 of 1,000 BRISC test images also appear in the Nickparvar dataset.

Frozen CLIP lightweight adaptation pipeline diagram

Human Activity Recognition Under Low-Light Conditions Using Frozen CLIP with Lightweight Adaptation

Md Ashik Khan, Abu Saleh Musa Miah, Fahmid Al Farid, Md Abdur Rahim, Hezerul Abdul Karim

Discover Artificial Intelligence (Springer Nature), vol. 6, art. 1006, 2026 · doi: 10.1007/s44163-026-02020-6

Frozen CLIP embeddings with lightweight adapters for activity recognition under low-light domain shift. Temporal adapters with 88.8K parameters raise ARID v1.5 accuracy from 68.37% to 74.38% (+6.01 pp), and every adapter adds under 0.2% FLOPs over the frozen backbone.

CNN-ViT hybrid methodology diagram

Fixed-Threshold Evaluation of a Hybrid CNN–ViT for AI-Generated Image Detection Across Photos and Art

Md Ashik Khan, Arafat Alam Jion

2025 28th International Conference on Computer and Information Technology (ICCIT), Cox's Bazar, Bangladesh, 2025, pp. 3853–3858 · doi: 10.1109/ICCIT68739.2025.11491484

We introduce a fixed-threshold evaluation protocol that holds decision thresholds fixed across all post-processing transformations (JPEG compression, blur, downscaling), exposing a forensic-semantic spectrum: frequency-aided CNNs collapse under compression (93.33% → 61.49%) while ViTs degrade minimally (92.86% → 88.36%).

Fixed-budget parameter-efficient training methodology diagram

Fixed-Budget Parameter-Efficient Training with Frozen Encoders Improves Multimodal Chest X-Ray Classification

Md Ashik Khan, Md Nahid Siddique

2025 28th International Conference on Computer and Information Technology (ICCIT), Cox's Bazar, Bangladesh, 2025, pp. 2748–2753 · doi: 10.1109/ICCIT68739.2025.11491458

Under a fixed parameter budget (2.37M, 2.51% of total), PET variants (BitFit, LoRA, adapters) achieve AUROC 0.892–0.908, outperforming full fine-tuning (0.770 AUROC) which uses 94.3M trainable parameters — a 40× reduction.

Resource-Efficient CNN methodology diagram

Comparative Analysis of Resource-Efficient CNN Architectures for Brain Tumor Classification

Md Ashik Khan, Rafath Bin Zafar Auvee

2024 27th International Conference on Computer and Information Technology (ICCIT), Cox's Bazar, Bangladesh, 2024, pp. 639–644 · doi: 10.1109/ICCIT64611.2024.11021970

Cited by 44

Accurate brain tumor classification in MRI images is essential for timely diagnosis. Our custom CNN achieved 98.67% accuracy on the Br35H dataset and 99.62% on the Brain Tumor MRI Dataset, showing competitive performance against pre-trained ResNet18 and VGG16 at lower complexity.

Research

My research is in computer vision and machine learning, with a focus on semi-supervised learning, resource-efficient methods and trustworthy AI. I apply these mainly to video action recognition and medical imaging.

A Large-scale Study of Representation Learning and Benchmarking in Video Action Recognition

Jul 2022 – Apr 2023

IIT Kharagpur · Supervised by Dr. Abir Das

  • Comprehensive analysis of 2D/3D CNNs and Transformer-based architectures for video action recognition
  • Evaluated representation transfer learning across 14 target datasets using 6 SOTA architectures (SlowOnly, TimeSformer, SIFAR)
  • Established benchmarks for cross-domain transfer learning and identified optimal transfer strategies
PyTorch MMAction2 FFmpeg

Deep Learning-based Hidden Camera Detection using Synthetic Training Data

Spring 2024
  • Novel synthetic data generation to address real-world data scarcity in surveillance detection
  • Fine-tuned ResNet50 and YOLOv8 with custom augmentation strategies
PyTorch OpenCV

Context Specific Quote Recommendation from Historical Text

Autumn 2022

IIT Kharagpur

  • Context-based quote recommendation using DistilBERT, achieving 82.25% accuracy
  • Full fine-tuning for quote phrase learning on the Quotation POTUS dataset
Python PyTorch Hugging Face

Evidence Retrieval for Fact Verification

Autumn 2022

IIT Kharagpur

  • Classical IR methods for evidence retrieval using tf-idf and cosine similarity
Python NLTK Beautiful Soup

OCR Service

Mar 2024

FastAPI web application performing OCR, NER (spaCy), and Sentiment Analysis on uploaded images with sync/async endpoints.

FastAPI Python Redis Docker

Education

Master of Technology

Computer Science and Engineering

Indian Institute of Technology, Kharagpur

2021 – 2023
Thesis

A Large-scale Study of Representation Learning and the Benchmarking in Video Action Recognition

Supervisor: Dr. Abir Das

Key Coursework
Machine Learning Information Retrieval Artificial Intelligence Complex Networks Data Analytics Algorithm Design Scalable Data Mining

Bachelor of Science

Computer Science and Engineering

Bangladesh Army University of Science and Technology, Saidpur

2015 – 2019
Key Coursework
Object Oriented Programming Data Structures Algorithms Database Management Machine Learning Digital Image Processing

Teaching

Teaching Assistant

Department of CSE, IIT Kharagpur

Software Engineering

Spring 2022–23

Algorithms-I

Autumn 2022

Conducted tutorials, evaluated assignments, mentored a group of 20 students during lab sessions, and supervised multiple student groups during projects.

Experience

Jan 2026 – Present

Senior Software Engineer

Auptimate

  • Developing and maintaining an e-signature platform for creating, managing, tracking and e-signing documents
  • Developing an AI Agent to assist customers with legal document-related processes
May 2023 – Dec 2025

Software Engineer

Auptimate

  • Developed and maintained an e-signature platform for creating, managing, tracking and e-signing documents
  • Developed an AI Agent to assist customers with legal document-related processes
Mar 2020 – Aug 2021

Assistant Programmer

Hovata Technologies

  • Developed PetrolERP — comprehensive web application for petroleum industry business process automation
  • Led development of Hovata Parking System for real-time parking management and optimization
  • Built REST APIs with Lumen Micro Framework and UI components with React JS and Redux
Sep 2019 – Feb 2020

Junior Software Developer

MerinaSoft

  • Developed inventory management system using Laravel
  • Developed Android applications and handled client requirements and UI design

Achievements

ICCR Scholarship for M.Tech at IIT Kharagpur

Awarded under the Bangladesh Scholarship Scheme to pursue postgraduate studies at IIT Kharagpur (2021–2023)

NVIDIA DLI Certificate

Fundamentals of Accelerated Data Science with RAPIDS (December 2021)

Training for Mobile Application Developer

Skill Development for Mobile Game & Application, ICT Division, Bangladesh (2018)

Finalist — BAUST IEEE Idea Contest

Recognized as a finalist in the BAUST-IEEE idea contest 2018

Honorable Mention — Programming Contest

Inter-University Programming Contest, IUBAT, Dhaka (2016)

Board Talent Rank 35th

SSC Examination, Dhaka Board (2012)

Contact

Email: aasshhik98@gmail.com · ashik.khan@kgpian.iitkgp.ac.in
Profiles: Google Scholar · LinkedIn · GitHub · LeetCode

Email

aasshhik98@gmail.com

ashik.khan@kgpian.iitkgp.ac.in

Phone

+880 1796 103496

Location

Mirzapur-1940, Tangail

Dhaka, Bangladesh