Yeasin Arafat

AI/ML & Full Stack Software Engineer working on computer vision, OCR, language models and the web apps around them.

I build models that find edited pixels and read Bangla handwriting, and I spend just as much time on the data and the checks behind them.

Md. Yeasin Arafat in a suit at the NSU ECE capstone showcase
person 0.99
Dhaka, Bangladesh Open to AI and software roles

I like the part where the model gets checked.

Most of my projects start with a dataset and end with a hard look at the errors. For my research paper I built a 9,611-image forgery set with pixel masks, then scored predictions on how well they traced the edges, since overlap alone hides sloppy boundaries.

For BhashaLLM, my teammates and I hand-wrote more than 1,500 pages of Bangla so our OCR model had real handwriting to learn from, and I presented the work as a paper at IEEE IS'26 in Bulgaria. In DocuMind, I wrote a checker that flags an LLM answer when its sources don't back it up.

I'm also comfortable building everything around the model: FastAPI and Django backends, Next.js front ends, and once, Tetris on a 128 by 64 pixel OLED screen for a plant monitor.

Lately I've been building open-source tools. Reviva restores old photos and videos and keeps nothing once it's done. Everything-videoEditing is an agent skill that lets Claude Code, Codex or Gemini CLI edit a video from one sentence, Bangla captions included.

Final-year B.Sc. in Computer Science and Engineering at North South University. Graduating April 2027. Bengali is my first language and I work in English every day.

Numbers I can explain

  • 0.874

    IoU localizing Stable Diffusion inpainting, up from a 0.849 baseline

  • 28% 12%

    Character error rate on handwritten Bangla OCR after QLoRA fine-tuning

  • 99%+

    Outputs in correct Bangla script from Llama 3.2-11B

  • 93.25%

    Accuracy on 7 fabric defect classes with gradient boosting I wrote from scratch

  • 9 LLMs

    Benchmarked on Bangla tasks, all on one 16 GB consumer GPU

Projects and research

Case files

Three projects where I did the data work, the modeling and the evaluation.

A TGIF-2 test photo where Stable Diffusion 2 painted in an aeroplane, with the model's predicted mask over it, next to the ground-truth mask, the predicted mask, Grad-CAM maps for the baseline and for our model, and IoU 0.874, Dice 93.29% and baseline IoU 0.849

Research paper, 2026. Supervised by Dr. Riasat Khan, NSU.

Finding the pixels Stable Diffusion painted

Text-guided inpainting leaves almost no visible trace. I wanted to know whether a segmentation model could mark exactly which pixels were generated, and how close it could get to the true edge.

I curated 9,611 images with pixel-level masks and kept COCO categories separate across splits so the test set couldn't leak. After benchmarking 12 segmentation models, I extended the strongest one, a ConvNeXt-L U-Net, with noise and frequency streams, edge-aware supervision and contrastive learning.

IoU
0.8742 from 0.8489
Dice
93.29%
GradCAM to mask
0.41 to 0.79

PyTorch, ConvNeXt-L, U-Net, SRM and DCT streams, GradCAM

View code on GitHub
BhashaLLM web interface showing Bangla OCR output from a handwritten page

IEEE conference paper and senior design project, 2025 to 2026. Supervised by Dr. Mohammad Rashedur Rahman.

Teaching open models to read and write Bangla

Most open LLMs treat Bangla as an afterthought. The small ones drift into English or slip Hindi letters into Bangla sentences. On top of that, everything had to run on a single 16 GB gaming GPU.

We benchmarked 9 open models from 1.5B to 12B parameters using a blind human rating rubric, fine-tuned in 4-bit with QLoRA, and wrote our own handwriting dataset of 1,500+ pages with line-level transcriptions for the OCR model.

OCR error rate
28% to 12%
Correct script
99%+ Llama 3.2-11B
Script mixing found
up to 23% small models

PyTorch, Transformers, PEFT and QLoRA, Qwen-VL, Llama 3.2, FastAPI

Illustration of DocuMind answering a question about a tenancy agreement, citing two clauses and passing the grounding check with a score of 0.91

Personal project, 2026.

An assistant that has to show its evidence

A retrieval system can still make things up. I wanted every answer checked against the passages it was supposed to come from before anyone reads it.

DocuMind parses PDFs and scanned images, retrieves evidence with FAISS, writes an answer, and then runs a verifier that scores how well that answer is grounded. Each one comes back labeled valid, weak or hallucinated, with a confidence score.

Pipeline
parse, retrieve, answer, verify
Tested
parser, retriever, verifier

Python, FastAPI, FAISS, sentence-transformers, OpenAI API, Tesseract, Docker

View code on GitHub

Research publication

Publication

Where the BhashaLLM work ended up.

Conference paper, September 2026. Published in IEEE Xplore.

From senior design project to IEEE paper

I presented BhashaLLM at the 13th IEEE International Conference on Intelligent Systems in Varna, Bulgaria, on 3 September 2026. The paper is now published in IEEE Xplore.

The paper covers our nine-model Bangla benchmark, the script mixing we found in small models, and the QLoRA fine-tuning that cut handwritten OCR error from 28% to 12%, all on a single 16 GB GPU.

13th IEEE International Conference on Intelligent Systems (IS'26), Varna, Bulgaria

BhashaLLM: A QLoRA-Based Framework for Bangla Text Generation and Handwritten Character Recognition

S. A. Binaaf, M. Y. Arafat, A. S. Chaklader, R. M. Rahman

Presented 3 September 2026. Published in IEEE Xplore, DOI 10.1109/IS71416.2026.11727246.

Events and life outside code

Behind the scenes

The capstone showcase, regular days at NSU, and what I do when I log off.

30 April 2026, NSU Plaza

Capstone Project Showcase, Season 19

The NSU ACM Student Chapter hosted the Capstone Project Showcase and Innovation Challenge Season 19 on April 30, 2026, at the NSU Plaza Area. We had booth 23, with our poster on the board and BhashaLLM running live on a desktop in front of it.

I was there with my teammates Samiyeel Alim Binaaf and Atique Shahrier Chaklader, explaining the project to whoever stopped by. Our supervisor was Dr. Mohammad Rashedur Rahman.

The ECE capstone batch and faculty in front of the showcase banner. Photo by Apon Photography.
Our poster, with the OCR pipeline and the model comparison.
Samiyeel, Atique and me at booth 23, with the demo on screen.

Around campus

Giving a presentation in a GED course.
In class and paying attention.

At the gym

I love going to the gym. It's where I clear my head after a long day at the computer, and I like having a number each week that I can try to beat.

Mirror check between sets.
A set on the press machine, filmed in the mirror.

Trying new restaurants

My other hobby is exploring places to eat. I go for the food, but a restaurant with a garden and warm lights, like this one, gets extra points from me.

On the garden steps after dark.
A seat by the plants and string lights.

More projects

More from the reel

More builds across ML, web apps, AI agent tools and hardware.

Chapter 01 2 projects

AI/ML & Computer Vision

Models that look at pictures: broken weaves in cloth, faces in old photos.

  • Fabric defect detection
  • Reviva photo and video enhancer
  1. Illustration of the fabric defect classifier: four fabric swatches with defect boxes, a 7-class confusion matrix, 93.25% accuracy and 23 ms per image

    2025, v2 in progress

    Fabric defect detection

    Gradient boosting written from scratch in NumPy, no scikit-learn. 93.25% accuracy across 7 defect types, 0.99 macro AUC and 23 ms inference.

    Version 2 is SpecTexNet, a CNN with a spectral branch that looks for the spots where the weave breaks. I found duplicate images in the v1 data, so v2 is tested under a leakage-free 5-fold protocol.

    Python, NumPy, ResNet50, Streamlit, PyTorch, ConvNeXt, Gradio

  2. Reviva web app: the upload screen with Photo and Video modes, a Restore faces toggle and a note that nothing is stored and no sign-in is needed

    2026

    Reviva photo and video enhancer

    Upload an old or blurry photo, or a short clip, and get a sharper version back. GFPGAN restores faces and Real-ESRGAN upscales. There are no accounts and no database, and files stay in RAM only until the result is sent.

    FastAPI, PyTorch, GFPGAN, Real-ESRGAN, FFmpeg, React, Vite, Tailwind, Docker

Chapter 02 2 projects

LLMs, RAG & Document AI

Language models that read documents and have to back up what they say.

  • AI Social Contract Advisor
  • DocuMind in the case files
  1. Illustration of the contract advisor: clauses rated high, medium and low risk, with the high-risk clause translated into Bengali

    2025

    AI Social Contract Advisor

    Reads contracts from messy PDFs and scans, rates each clause as low, medium or high risk against a scoring rubric, and translates the result into Bengali.

    Next.js, Django, PostgreSQL, Gemini API, Tesseract

    Code

Chapter 03 2 projects

Full Stack Web Applications

Whole apps: the interface, the API, the database and the AI inside.

  • AI e-commerce store
  • SweatFit AI
  1. Illustration of the store admin: a plain-language product description turned into a full listing with a suggested price, next to a live customer chat

    2026

    AI e-commerce store

    Admins describe a product in plain words and GPT-4o turns it into a full listing with pricing ideas. Customers and admins chat live over Socket.io.

    Next.js 15, TypeScript, MongoDB, Socket.io, Clerk

    Code
  2. Illustration of the SweatFit dashboard: daily calories, macros, workouts this week and the plan for today

    2024

    SweatFit AI

    Workout, nutrition and progress tracking with calorie and meal APIs, JWT login and a responsive dashboard. My first full-stack AI app.

    React, Django, PostgreSQL

    Code

Chapter 04 2 projects

AI Agents & Developer Tools

Skills that make coding agents do more and check their own work.

  • Everything-videoEditing
  • Loopify
  1. Everything-videoEditing demo: one tea brand ad exported as a 9:16 Reel, a 16:9 YouTube video with subtitles and a 4:5 feed post

    2026

    Everything-videoEditing

    An agent skill that turns Claude Code, Codex or Gemini CLI into a video editor. It studies the footage, asks two to four quick questions, then cuts, grades, captions and exports Reels, YouTube and feed versions. Bangla captions render with correct shaping.

    Python, FFmpeg, Whisper, Demucs, CLAP, Remotion, Blender

    Code
  2. Loopify banner in pixel letters above the flow: plain task, verifier, work and check, verified result

    2026

    Loopify

    A prompt skill that turns a plain task into a loop. It writes the check first, keeps working until the check passes, then tells you how far to trust that check. One core file, with small adapters for Claude, ChatGPT, Gemini and Grok.

    Markdown, Shell, JavaScript

    Code

Chapter 05 1 project

Embedded Systems

Firmware in C++ on a small STM32 board, with real sensors.

  • PLANT-X plant monitor
  1. PLANT-X prototype on a breadboard with OLED screen and keypad

    2026

    PLANT-X plant monitor

    An STM32 board that watches soil, heat, humidity and light, beeps when a plant needs help, and hides five games on a tiny OLED. Built with a team of four.

    C++, STM32, PlatformIO, I2C, ADC

    Code

Technical skills

Loadout

What I reach for, grouped by the job.

I work on the whole path from data to the screen. I clean and label the data, train and check the model, serve it through an API, and build the interface people use.

Datalabels, splits Modeltrain, evaluate APIFastAPI, Django InterfaceReact, Next.js
Full stack AIFrom raw data to a screen people use

Vision

PyTorch, TensorFlow, OpenCV, Albumentations, semantic segmentation, U-Net, ConvNeXt, GradCAM, GFPGAN, Real-ESRGAN

Language

Hugging Face Transformers, PEFT, QLoRA and LoRA, BitsAndBytes, RAG, FAISS, sentence-transformers, Tesseract, PyMuPDF

Data and evaluation

Dataset curation, pixel masks, bounding boxes, leakage-free splits, IoU, Dice, Boundary F1, Hausdorff95, CER, ROUGE, confusion matrices, ROC and AUC, human rating rubrics

Building

Python, TypeScript, JavaScript, C++, C, Java, SQL, FastAPI, Django REST, Next.js, React, Vite, Tailwind CSS, Gradio, PostgreSQL, MongoDB, Docker, Git, GitHub Actions, Linux, Hugging Face Spaces

Hardware

STM32, PlatformIO, GPIO, ADC, I2C, keypad scanning, SWD debugging

Media and AI agents

FFmpeg, Whisper, Demucs, CLAP, Remotion, Blender, edge-tts, and agent skills for Claude Code, OpenAI Codex and Gemini CLI

Education and timeline

How I got here

  1. 2018

    Secondary School Certificate

    Science, Model Academy. GPA 5.00 out of 5.00.

  2. 2020

    Higher Secondary Certificate

    Science, Bangladesh Navy School and College. GPA 5.00 out of 5.00.

  3. 2022

    Started CSE at North South University

    Machine learning, algorithms, databases, operating systems and networks.

  4. 2025

    BhashaLLM begins

    Senior design project on Bangla LLMs and handwritten OCR.

  5. 2026

    Forgery research, DocuMind, PLANT-X

    The year I got serious about evaluation.

  6. Apr 2026

    Showed BhashaLLM at the Capstone Project Showcase

    Booth 23 at Season 19, hosted by the NSU ACM Student Chapter.

  7. Sep 2026

    Presented BhashaLLM at IEEE IS'26

    13th IEEE International Conference on Intelligent Systems, Varna, Bulgaria.

  8. Oct 2026

    BhashaLLM published in IEEE Xplore

    DOI 10.1109/IS71416.2026.11727246.

  9. Oct 2026

    Reviva, Everything-videoEditing and Loopify

    Open-source tools, plus version 2 of the fabric defect project.

  10. Apr 2027

    Graduation

    Looking for my first full-time role in AI or software engineering.

Credits

Written, trained and debugged by
Md. Yeasin Arafat
Pixel masks
9,611 images, checked by hand
Handwriting
The BhashaLLM team, 1,500+ pages
Research supervision
Dr. Riasat Khan
Senior design supervision
Dr. Mohammad Rashedur Rahman
Premiered at
IEEE IS'26, Varna, Bulgaria
Published in
IEEE Xplore
Hardware
PLANT-X team, CSE 331L Group 9
Side projects
Reviva, Everything-videoEditing, Loopify
Compute
One RTX 5070 Ti, 16 GB, doing its best
Filmed on location
Dhaka, Bangladesh

Audience reviews

What people say

    No reviews yet. Yours could be the first.

    Leave a review

    Your name and email will be shown with your review.

    Rate my portfolio
    Relationship with me

    Box office

    Who's been watching

    – watching now

    –

    Total views

    –

    Views in the last 7 days

    Views per day, last 7 days

    IEEE IS'26 certificate of attendance for Md. Yeasin Arafat