I work on computer vision that holds up outside the lab: on crowded roads, in classrooms and in humanitarian settings.
I give what I learn away for free.
I am looking for a fully funded PhD in AI and for sponsors. Every sponsored project is released open for all.

• 🎯 Seeking: a fully funded PhD in AI for Fall 2027, in computer vision and medical image analysis.
• 🔬 Research focus: robust, label-efficient and well-calibrated vision models for medical images (blood-cell microscopy, polyps, skin lesions) that keep working under acquisition and stain shift and run on modest hardware.
• 📝 Research record:
  • 1 peer-reviewed paper, published by Springer Nature (2026): Emotion Detection at Lightning Speed: Leveraging YOLOv11 for Facial Expression Analysis, Lecture Notes in Networks and Systems (ICCTRDA 2025).
  • 1 paper under review: a three-seed benchmark of nano YOLO detectors for blood-cell detection, submitted to IEEE ISBI 2027 (with H. J. Shethil and J. A. Aranna).
  • 6 manuscripts in preparation: blood-cell classification, calibration and label audit (target IEEE EMBC 2027); stain-robust, label-efficient blood-cell detection (target MICCAI/MIUA 2027); polyp segmentation with YOLO and SAM 2 (target MIUA 2027); zero- and few-shot medical image classification with BiomedCLIP (target MICCAI 2027 workshop); CPU deployment of blood-cell classifiers in Python and C++; people counting in 4K drone video (target Machine Vision and Applications).
  • 3 undergraduate theses: pedestrian risk, vehicle over-speed and brain-tumor classification.
• 💼 Now: Advanced ICT Officer at CNRS-UNHCR since October 2025. I teach students digital and ICT skills in a humanitarian setting.
• 🧪 Before: Research Assistant at Robo Tech Valley, Dhaka (March–July 2025), working on AI/ML and computer-vision projects.
• 🏛️ Education: B.Sc. in Computer Science & Engineering, Ahsanullah University of Science and Technology (AUST), 2025.
• 🌍 Open science:
  • 273 free written lectures and 110 free animated video lectures;
  • 100 free GRE and TOEFL study apps, on the web and for Windows;
  • 100 free IELTS study apps, on the web and for Windows;
  • every sponsored project is released open for all.
• ⚡ Algorithms: 12+ years of competitive programming (since 2014).

I want to build vision models that keep working when conditions change: a new microscope, a different stain, a darker camera or a cheaper computer.

My undergraduate research tackled pedestrian detection with risk-level classification for Bangladeshi road scenes, and vision-based vehicle over-speed detection. My paper on real-time facial-expression analysis with YOLOv11 was published by Springer Nature in 2026. Since then I have focused on medical image analysis: a multi-seed benchmark of nano YOLO detectors for blood-cell detection (under review at IEEE ISBI 2027), classifier failure under acquisition and stain shift, few-shot learning with BiomedCLIP, CLIP and DINOv2, and CellShift, label-free adaptation to a new laboratory with calibrated cell counts.

In a PhD I want to work on robust, label-efficient and well-calibrated medical imaging models that can run on modest hardware in low-resource clinics.

A. Giri, J. A. Apurba, A. Z. M. J. Uddin, S. I. Rizon, M. R. Begum, A. S. M. A. S. Akib, M. Giri, B. Shahi, M. Khadgi. “Emotion Detection at Lightning Speed: Leveraging YOLOv11 for Facial Expression Analysis.” In Proceedings of the International Conference on Computational Technologies for Research in Data Analytics (ICCTRDA 2025), Lecture Notes in Networks and Systems, Springer Nature, 2026, pp. 218–233. DOI: 10.1007/978-3-032-27666-7_19

🎓 Research profiles: Google Scholar · ORCID · ResearchGate · Academia.edu

1. A new lab, no new labels: When a detector trained in one laboratory meets a new microscope or stain, how much accuracy can be recovered without labels, and how do we know when the counts can be trusted?
2. Few labels, big models: When do foundation models such as BiomedCLIP beat a small supervised model, and how few labelled images does a clinic really need?
3. Beyond the benchmark: How can detectors stay reliable on dense, unstructured roads, with rickshaws, buses and pedestrians sharing a lane in rain or at night, when they were trained on data from elsewhere?
4. Real-time on modest hardware: How far can YOLO-class models be compressed for edge devices without losing accuracy on the cases that matter most for safety?
5. Fair human understanding: Do expression and pose models work equally well across ages, skin tones and cultures? How do we measure the gaps and close them?
6. AI for humanitarian settings: How can vision and learning systems responsibly support education and services in refugee and low-resource contexts?

I build in the open, but compute, data and publishing cost real money.

My pledge: when a project is sponsored, I release it openly and free for anyone to learn from, use and build on. That includes:

• the source code;
• the trained model weights;
• the datasets, with personal data anonymised and consent respected;
• an open-access paper;
• lessons on my free teaching sites.

What sponsorship pays for

• 🖥️ GPU compute to train and evaluate vision models
• 🏷️ Collecting and annotating real-world data, such as Bangladeshi road scenes
• 📄 Open-access publication fees, so the papers are free to read
• 🎬 More free lectures and animated lessons
• ✈️ Travel to present the work at conferences

Projects open for sponsorship

Sponsors get:

• acknowledgement in the repository, papers and lectures;
• regular progress updates;
• a direct line to me throughout the project.

• The AI Lecture Hall · source
  • 273 university-level lectures in 10 tracks.
  • Covers AI foundations and mathematics through transformers, diffusion models, reinforcement learning, MLOps and AI ethics.
  • Each lecture includes derivations, code and exercises.
• AI in Motion · source
  • 110 animated video lectures: more than 8 hours, including 36 deep dives of 10+ minutes.
  • Narration, subtitles, chapters, quizzes and printable illustrated notes.
  • Every animation is computed live from the real algorithm: gradient descent really descends and Q-learning really learns.
• Machine Learning: Basics to Intermediate, in Bangla · on JaniX, my learning platform
  • 4 video lessons (about 1 h 48 min) in Bangla with English subtitles, from first principles to modern AI.
  • Every idea is worked out by hand with real numbers, then coded in Python (scikit-learn, PyTorch).
  • Each part ends with a quiz and homework; passing the quizzes earns a certificate.
• GRE and TOEFL Study Suite · source
  • 100 free study apps for the GRE and the January 2026 TOEFL iBT, in the browser or as Windows programs, with no account.
  • Vocabulary, generated quant questions with exact answers, verbal and reading practice, listening read aloud, and speaking with voice recording.
  • Timed sections, mock tests, an error log and a study planner.
• IELTS Study Suite · source
  • 100 free study apps for IELTS Academic and General Training, in the browser or as Windows programs, with no account.
  • Generated Part 1 calls, maps and Task 1 charts; original listening scripts and reading passages with every question type.
  • Essay and letter practice, a Speaking test recorded on your device, vocabulary, band conversion and mock tests.
• Teaching at CNRS-UNHCR: ICT and digital skills for students.
• YouTube: Study with Janin and Pomodoro Study with Janin.

• 12+ years of contest problem solving, since 2014, on Codeforces, CodeChef, HackerRank and other platforms.
• Recognised for outstanding performance in online programming contests.
• 📈 Profile: StopStalk
• 📂 Archives:
  • Competitive-programming resources collection
  • Codeforces solutions

Area	Tools
AI and machine learning
Algorithms and systems
Web and backend
Data
Tools

• Advanced Learning Algorithms · DeepLearning.AI & Stanford Online
• Unsupervised Learning, Recommenders, Reinforcement Learning · DeepLearning.AI & Stanford Online
• Divide and Conquer, Sorting and Searching, and Randomized Algorithms · Stanford University
• Advanced Algorithms and Complexity · UC San Diego
• Object-Oriented Data Structures in C++ · University of Illinois Urbana-Champaign
• Data Scientist Career Track · 365 Data Science: exams in ML, deep learning, SQL, mathematics, probability and statistics
• Convolutional Neural Networks with TensorFlow in Python · 365 Data Science
• MLOps Bootcamp · DataMites
• Introduction to DevOps · IBM

📁 All 39 certificates (Google Drive)

• 🥇 Talent Pool Scholarship (2008): ranked 4th in Kurigram District.
• 🏆 Competitive-programming awards for performance in online contests.
• 🧠 Microsoft MVP Program trainee, Dhaka: workshop on AI, computer vision and Microsoft Azure.
• 🔶 Google Developer Group TensorFlow workshop, Dhaka: AI, machine learning and data mining.

• 📧 Email: japurba97@gmail.com
• 🌐 Academic portfolio: janin-a-apurba.vercel.app
• 🎓 Google Scholar: Janin A Apurba · ORCID: 0009-0002-8576-284X
• 🔬 ResearchGate: Janin-Apurba · Academia.edu: JaninAApurba
• 📄 CV: Google Drive
• 👥 References: academic references from AUST CSE faculty are available on request.

If my work helped you, please follow and subscribe:

• Study with Janin: youtube.com/@studywithjanin
• Pomodoro Study with Janin: youtube.com/@pomodorostudywithjanin3326

_{© Janin A Apurba, CSE, AUST · Advanced ICT Officer, CNRS-UNHCR}