Sakib on a mountain, the ridges behind him traced in red code

Tauhidur Rahman Sakib — Machine Learning Engineer & AI Researcher

I teach machines to see, read and decide.

Machine Learning Engineer, AI developer and mountaineer. I take on hard problems the way I take on peaks: one steady step at a time, all the way to the top.

See My Work Get in Touch

About

Give me the problem. I’ll build the AI model that solves it.

I’m Tauhid, a Machine Learning researcher and engineer focused on advancing AI through rigorous, high-impact work. I own the full research pipeline: problem framing, data, model design, experimentation, validation and publication. That work has produced a portfolio of peer-reviewed papers in machine learning and deep learning. I’ve led as both first author and research supervisor, setting direction, coordinating collaborators and taking projects from idea to accepted paper. That’s where my strengths in research leadership, technical writing and problem-solving come from. I don’t limit myself to one domain. Beyond research I build across the stack, from software development and automation to AI engineering, system design and emerging technologies. I learn fast, take on unfamiliar problems, and compete at a high level. I get my hands dirty in every corner of AI. Right now I’m building at the intersection of AI and blockchain: AI-driven trading agents that execute with precision and create a trading experience people can trust. Research and backtesting are underway, paper-trading results are very strong, and the product launches soon. Open to research collaborations, ambitious technical challenges, and roles across the IT industry where curiosity, adaptability and strong execution make the difference.

Experience

  1. Present

    Research Assistant, Machine Learning

    Chang’an University, School of Information Engineering · Remote

    Transformer model-editing research with Prof. Zhaoyang Zhang using the COAR framework. Analyzed 1,224 DistilBERT components; contrastive ranking reached selectivity ratios of 2.36–27.15 (vs 0.81–1.24) with almost no accuracy loss. Manuscript under review at a Q1 journal.

  2. Industry

    AI Developer

    Devartistry

    Designed and deployed NLP chatbot assistants for client websites to automate customer support, and integrated conversational AI into web applications.

  3. Research

    Research Supervisor

    Independent ML research projects

    Led multiple ML/DL papers from study design and data preparation through modeling, validation and writing, coordinating collaborators as first author.

  4. 2020 — 2025

    B.Sc. in Computer Science and Engineering

    BRAC University · Dhaka

    Focused on machine learning, deep learning and research.

Selected Work

01AI × BlockchainComing soon

Freyr

Closed source · Subscription-based

Freyr is an AI trading agent for crypto markets. It watches live market data, decides when to enter and exit a trade, and places the orders itself. Every decision is recorded on the blockchain, so its track record is transparent and can’t be edited after the fact.

  • Reads live price, volume and trend data and scores each trade opportunity in real time
  • Executes buy and sell orders automatically, with built-in risk limits
  • Logs every action on-chain for a tamper-proof, verifiable history
  • Research and backtesting are underway; paper trading is showing very strong results

AI agents / Blockchain / Algorithmic trading / Backtesting

02Research · 2026

COAR Model Editing

Language models are black boxes: when one gets something wrong, it’s hard to tell which internal part caused it. This research uses COAR (Component Modeling and Editing via Random Ablations) to pinpoint the exact components inside a transformer that drive a behaviour, then edits only those, changing that behaviour without breaking anything else.

  • Switches off random groups of components thousands of times and measures how the output changes
  • Scores all 1,224 DistilBERT components by their real influence on the prediction
  • A new contrastive ranking was up to 27× more selective than the standard method (2.36–27.15× vs 0.81–1.24×)
  • Editing the top components changed held-out accuracy by at most 0.0006
  • With Prof. Zhaoyang Zhang, Chang’an University · under review at a Q1 journal

Transformers / Interpretability / Model editing / DistilBERT

View on GitHub

03AI tool · 2026

Backbencher AI Copilot

Free to use · No API key · Runs locally

A real-time AI copilot for interviews and viva exams. It listens to the interviewer through the computer’s audio, transcribes the question on the device, and shows a clear, structured answer in a private overlay while the conversation is still going.

  • Stealth mode: the overlay is invisible to screen sharing and recording. Even when the candidate shares their whole screen, nobody else sees it; only the person using it can.
  • Captures system audio, so it works with Google Meet, Zoom and Teams
  • Free, no API key needed: it runs on local LLMs, so there are no subscriptions, no usage limits and nothing is sent to the cloud
  • With a strong GPU, the whole pipeline (transcription and answers) runs entirely on your own machine
  • Transcribes speech locally on the device for speed and privacy
  • Turns each question into short key points to speak from, within seconds

Local LLMs / Speech-to-text / Stealth overlay / Real-time

View on GitHub

Research

I have authored 15 research papers: 11 published with IEEE, Elsevier, IET and SSRN, and 4 more recently accepted and coming online soon. I am first author on 3, supervisor and corresponding author on 10, and a top contributor on 4 more. Most of my work is in medical imaging AI, building models that diagnose diabetic retinopathy, retinal disease, lumbar spinal stenosis, Alzheimer’s disease, brain stroke, rare neurological disorders, liver fibrosis and colon polyps, often with hybrid CNN and Vision Transformer architectures validated on external clinical data. Beyond medicine, I research transformer interpretability and model editing, NLP for spam detection, drug-mechanism prediction, industrial defect inspection, defence reliability and data-driven mental-health studies.

SSRN2026
SupervisorCorresponding author
SSRNPublished
A Lightweight Network Matches Larger Models in Cross-Dataset Polyp Segmentation under a Controlled Multi-Seed Protocol

Abstract—Cross-dataset generalization and deployment cost are two largely independent obstacles to the clinical translation of polyp segmentation. We present WaveLite-DG, a compact encoder–decoder addressing both. It combines a MobileNetV3-Large encoder, Wavelet Frequency Modules that perturb approximation-band feature statistics during training while enhancing detail-band structure, a gated context scan at the bottleneck, and a boundary-supervised decoder. We derive in closed form the energy this perturbation injects. Trained on Kvasir-SEG and CVC-ClinicDB, with CVC-ColonDB, ETIS-LaribPolypDB and CVC-300 withheld from every fitting and selection decision, WaveLite-DG reaches a three-seed mean macro unseen Dice of 0.8300 ± 0.0004, the highest of six architectures retrained under one protocol without photometric augmentation and statistically equivalent to Polyp-PVT and VM-UNet within a pre-set margin of 0.02. With colour jitter, or with the perturbation added to those two baselines, they lead by under 0.008. In every regime tested WaveLite-DG stays within 0.008 of the best model with one third or less of the parameters (7.68 M) and one fifth or less of the computation (1.93 giga multiply–accumulate operations), and it is 0.0427 more accurate than the best of four lightweight networks. Approximation-band perturbation adds +0.0651 macro Dice without photometric augmentation but only +0.0185 with colour jitter, which recovers most of the gain, and perturbing the undecomposed feature at matched energy shows no detectable difference. The contribution therefore lies in efficiency at parity rather than in a new mechanism.

Index Terms—polyp segmentation, domain generalization, wavelet transform, lightweight networks.

I. Introduction

Colonoscopy models often fail on hospitals they were not trained on, and large models are costly to deploy.

Fig. 1.
II. Related Work
III. Method
IV. Results
SSRN preprint

A Lightweight Network Matches Larger Models in Cross-Dataset Polyp Segmentation under a Controlled Multi-Seed Protocol

SSRN preprint

Tap or swipe to read the abstract

Google Scholar profile

Certificates

Recognised on stage, not just on paper.

My research has been recognised beyond publication. At ICCSDFAI 2026 in Istanbul, Türkiye, our lumbar spinal stenosis paper won the 1st-place Best Paper Award, chosen as the best work presented at the conference. I have also presented papers at international conferences: two at ICCSDFAI 2026 (spinal stenosis and diabetic retinopathy), three at the 5th IEEE SPICSCON 2026 in Bangladesh (rare neurological disease, liver fibrosis and steel defect detection), and one at iCONEECT 2026 at Premier University, Chattogram (brain-stroke classification). Outside research, BRAC University’s Art & Photography Society awarded me a Certificate of Achievement for leadership and promoted me to Secretary of its Finance & Event Management department.

Best Paper Award · 1st Place, ICCSDFAI 2026 · Istanbul, Türkiye Award2026 Best Paper Award · 1st Place ICCSDFAI 2026 · Istanbul, Türkiye For “A Hybrid Dual-Backbone Deep Learning Framework for Lumbar Spinal Stenosis Classification from MRI with External Validation” View certificate

Services

If you can describe it, I can build it.

I take projects end to end, from idea to deployed product. AI models and computer vision, LLM chatbots and agents, full-stack web apps, automation, trading tools and research support, with clear scope and transparent pricing.

  • 01AI & Machine Learning
  • 02LLM Apps & AI Agents
  • 03Web & App Development
  • 04Automation & Data
  • 05Blockchain & Trading Tech
  • 06Research & Academic Support
  • 07Mobile App Development
  • 08UI/UX & Product Design
  • 09Generative AI & Content
  • 10NLP & Text Analytics
  • 11MLOps & Cloud Deployment
  • 12E-commerce & Business Systems
  • 13SaaS & Startup MVPs
  • 14Data Engineering & Databases
  • 15AI Consulting & Training
  • 16Cyber Security & Ethical Hacking

Recommendations

Vouched for by the people I’ve worked with.

Two professors who have supervised my work closely have written letters of recommendation for me: my undergraduate thesis supervisor at BRAC University, and the professor I work with on transformer model-editing research at Chang’an University. Copies are available on request.

Letter of Recommendation

Dr. Muhammad Iqbal Hossain

Associate Professor
BRAC University, Dhaka

Undergraduate thesis supervisor

Available on request

Letter of Recommendation

Prof. Zhaoyang Zhang

Professor, School of Information Engineering
Chang’an University, China

Research supervisor · transformer model editing (COAR)

Available on request
Request a letter

Tools

Languages
Python, JavaScript, HTML/CSS
Machine Learning
XGBoost, Random Forest, Gradient Boosting, stacking, SVM, autoencoders, SMOTE, K-fold CV
Deep Learning & NLP
CNNs, Vision Transformers, BERT, DistilBERT, transfer learning, segmentation, Grad-CAM, chatbots
Data & Stats
EDA, feature engineering, visualization, SPSS, SmartPLS, ANOVA, hypothesis testing
Workflow
Jupyter, Git, GitHub, CI/CD, Docker, LaTeX

Human ideas. Written in code.

Every product starts with a person and a problem. I write the code that connects them.

I work in AI development: from medical-imaging models to NLP chatbots, I build AI that solves real problems.

FAQ

Who is Tauhidur Rahman Sakib?

Md. Tauhidur Rahman Sakib is a Machine Learning Engineer and AI researcher from Dhaka, Bangladesh, with a B.Sc. in Computer Science and Engineering from BRAC University. He has authored 15 research papers (IEEE, Elsevier, IET, SSRN) and won a 1st-place Best Paper Award at ICCSDFAI 2026.

What does Tauhidur Rahman Sakib research?

His research focuses on medical imaging AI (diabetic retinopathy, spinal stenosis, Alzheimer’s, brain stroke, liver fibrosis, polyp segmentation), hybrid CNN and Vision Transformer models, transformer interpretability and model editing (COAR), NLP and applied machine learning.

What services does he offer?

He offers 60 services across 16 categories, including custom machine learning models, computer vision, LLM chatbots and AI agents, full-stack web and mobile apps, automation, data engineering, cyber security and ethical hacking, blockchain and trading tools, and research support, with transparent USD price ranges.

How can I hire Tauhidur Rahman Sakib?

Message him on WhatsApp (+880 1776 606340), email tauhidurrahmansakib@gmail.com, or add services to the cart on the services page and send the order on WhatsApp. Every project starts with a free scoping call and a fixed quote.

Is he open to research collaborations?

Yes. He is open to research collaborations, supervision roles and ambitious technical projects, especially in medical imaging AI, transformers and applied machine learning.

Support

Contact

Have a hard problem? Let’s climb it together.

tauhidurrahmansakib@gmail.com

Dhaka, Bangladesh · Open to remote