Skip to content
Ajoy Sarker

PhD Applicant — Fall 2026 / 2027

Ajoy Sarker

Machine Learning · Deep Learning · Signal Processing · NLP

I want to do a PhD on robust deep learning for biomedical signals and low-resource NLP—building models that hold up on noisy, real-world data and can run efficiently at the edge.

Currently building production AI as an AI/ML Engineer at Brain Station 23, while pursuing research in EEG-based emotion recognition and low-resource Bengali NLP.

Download CVEmail meScholarLinkedInGitHub
Research output
3papers
Industry ML
2+yrs
M.Sc. CGPA
3.75/ 4.00
Production scale
100K+/ day
01

About

EEG signal processing · Deep learning · Efficient AI · Low-resource NLP

AI/ML researcher and engineer working at the intersection of biomedical signals, deep learning, and low-resource NLP.

My research focuses on robust models for noisy, real-world data—EEG and other biomedical time-series, and Bengali NLP under heavy class imbalance and code-switching. I care about end-to-end pipelines: from preprocessing and feature design to deep learning architectures (CNNs, RNNs, Transformers) and rigorous evaluation.

In parallel, I build production AI at Brain Station 23: multimodal pipelines (vision, OCR, audio), retrieval systems with sub-200 ms latency, and inference optimization for edge-style deployment. That keeps me honest about what works once the lab notebook is closed.

I am especially interested in temporal modeling, multimodal learning, and edge-efficient architectures for medical signals/imaging and environmental sensing.

Open to PhD positions starting Fall 2026 or Fall 2027—happy to discuss potential collaborations or fit.

02

Education

  1. Jahangirnagar University

    M.Sc. in Computer Science & Engineering (AI / ML)

    May 2024 – Sept 2025CGPA · 3.75 / 4.00

    Thesis direction: EEG-based emotion recognition with deep learning—preprocessing, band-power features, and CNN/RNN/Transformer models.

    Dhaka, Bangladesh

  2. Jahangirnagar University

    B.Sc. in Computer Science & Engineering

    Feb 2019 – Mar 2024CGPA · 3.68 / 4.00

    Thesis: Bengali abusive language detection—stacked ensembles with multi-level TF–IDF features, +1.9% over prior SOTA.

    Dhaka, Bangladesh

03

Research

Robust ML on noisy real-world data—biomedical signals, low-resource NLP, and large tabular records.

Unpublished

EEG-Based Emotion Detection

Problem

Inferring emotion from EEG is a hard ML problem: non-stationary signals, artifacts, and long-range temporal structure across multichannel data.

Approach

  • Signal preprocessing — bandpass filtering, artifact removal, ICA
  • Band-power features across delta, theta, alpha, beta, gamma
  • Deep learning for sequences: CNN, RNN, and Transformer baselines
Datasets
DEAP, MAHNOB-HCI
Stack
TensorFlow · NumPy · SciPy
Goal
Robust multichannel EEG emotion classification
Under Review

Abusive Language Detection in Bengali

Problem

NLP for a low-resource language: severe class imbalance, rich morphology, code-switching, and culturally grounded abuse—harder than off-the-shelf English models.

Approach

  • Multi-level TF–IDF features (word + character n-grams) and linguistic signals
  • Stacked ensemble: LinearSVC, logistic regression, XGBoost
  • Enhanced BD-SHS corpus with 15-class taxonomy (~58k samples)
Accuracy
91.2% (15-class)
Improvement
+1.9% over prior SOTA
Stack
scikit-learn · XGBoost
BUET-ARI Collaboration

Road Accident Severity Prediction

Problem

Tabular ML at scale: predicting crash occurrence and severity from imbalanced, noisy administrative records covering all 64 districts of Bangladesh.

Approach

  • Gradient-boosted trees and ensembles: CatBoost, XGBoost, Random Forest
  • Feature engineering across demographic, environmental, and infrastructure fields
  • Class imbalance handling via SMOTE and cost-sensitive learning
Dataset
500K+ records · 2001–2020
F1 Score
0.65
Stack
CatBoost · XGBoost · scikit-learn
04

Publications

  • Abusive Language Detection in Bengali Social Media

    Ajoy Sarker, M. Z. Rahman

    Under review · 2025

    Read now

    NLP / text classification for Bengali social media: a stacked ensemble (LinearSVC, logistic regression, XGBoost) over multi-level TF–IDF features on an enhanced 15-class BD-SHS corpus. Reaches 91.2% accuracy and 90.7% macro-F1—about +1.9 points over prior work—with strong per-class scores and inference suitable for near–real-time moderation.

    Key contributions

    • Features tuned to Bengali morphology, code-switching, and cultural context—not generic English pipelines.
    • Statistically significant lift over earlier Bengali baselines (p < 0.001) via stacking and tuning.
    • Extended BD-SHS to 15 categories (~58k samples) with category-wise metrics for low-resource moderation.

  • Road Accident Severity Analysis and Prediction

    Ajoy Sarker, S. N. Riya, U. F. Moon, M. A. H. Rasel, M. M. Uddin, Md M. Anwar

    Working paper · BUET-ARI

    Read now

    Applied ML on tabular crash data: predicts accident occurrence and severity at district scale from merged ARI/BUET police records and BRTC data (2001–2020, all 64 districts), with cleaning, imputation, and feature selection for large, noisy administrative tables.

    Key contributions

    • One of Bangladesh's largest longitudinal crash datasets used end-to-end for supervised ML.
    • Gradient boosting and tree ensembles—CatBoost leads occurrence (F1 ≈ 0.65); gradient boosting best for severity (R² ≈ 0.082 tuned).
    • Findings framed for transportation authorities: where crashes concentrate and how severely.

  • Enhancing Emotion Recognition from EEG Signals

    Ajoy Sarker, S. N. Riya, Imdadul Islam

    Manuscript · 2024

    Read now

    ML on high-dimensional EEG: three-way valence (positive / negative / neutral) from Muse data—1,982 trials, 2,549 statistical and band features per trial—and a benchmark of 11 models from linear baselines to kernel SVMs, boosting, and wide ANNs.

    Key contributions

    • Compares classical ML against non-linear and neural models—kernel SVMs and wide ANNs reach up to ~98.4% test accuracy in reported runs.
    • Argues feature-rich representations are key to separating overlapping EEG classes; analyzes why purely linear models break down.
    • Documents tuned kernels, ANN widths, and ensemble options, with implications for HCI and real-time affective computing.
05

Open Source Models

Public Hugging Face checkpoints—training pipelines and evaluation built end-to-end.

Bangla Math LLMs (Gemma & LLaMA)

Fine-tuned LLMs for Bengali mathematical reasoning. Built the dataset, training pipeline, and evaluation benchmarks end-to-end.

  • LoRA
  • PEFT
  • Bengali
  • Reasoning

Bangla Sarcasm Bi-Encoder

Dual-encoder model that improves semantic understanding for sarcasm detection in low-resource Bengali NLP.

  • Sentence-Transformers
  • Dual Encoder
  • Low-Resource NLP
06

Skills & Methods

Core ML / DL

  • Python
  • PyTorch
  • TensorFlow
  • scikit-learn
  • XGBoost
  • CatBoost
  • Signal processing (EEG, time-series)

ML Systems & NLP

  • Transformers (LLaMA, Gemma)
  • LoRA / PEFT fine-tuning
  • Sentence-Transformers
  • RAG · Semantic search
  • Multimodal pipelines (vision + OCR + audio)

MLOps & Infra

  • Docker
  • CI / CD
  • Prometheus
  • Grafana
  • AWS

Backend & Data

  • FastAPI
  • NestJS
  • PostgreSQL + pgvector
  • MongoDB
07

Selected Projects

Applied AI shipped in production—where research and engineering meet.

NLP · LLMs · Enterprise

Production RAG System

  • Retrieval-augmented generation with OpenAI and Gemini
  • Embeddings + pgvector for production-scale semantic search
  • Sub-200 ms latency, 92% retrieval accuracy
  • Deployed across 20+ enterprise customers

Vision · OCR · Audio

Multimodal Data Pipeline

  • Gemini Vision, Tesseract, EasyOCR, Deepgram
  • Fuses OCR, speech-to-text, and structured fields for downstream models
  • Processes 100K+ records daily
  • 94% end-to-end accuracy

Optimization · MLOps

Edge-Oriented Inference

  • Quantization and batching for ~50% faster inference
  • Latency- and resource-constrained deployment
  • Drift detection and monitoring (Prometheus, Grafana)

AI Hackathon · 11 / 90

Hybrid Search System

  • Dense + sparse retrieval (vectors + keywords)
  • Safety guardrails on LLM outputs
  • 50% faster inference via model and system optimization
08

Honors & Achievements

  • NST Fellowship — Ministry of Science & Technology, Bangladesh
  • University Scholarship — Awarded 4× for academic excellence
  • AI Hackathon 2025 — 11 / 90 teams
  • CodeSamurai 2022 — 29 / 500+ teams
  • Competitive Programming — 2,000+ problems solved
09

References

Available on request

Recommendation letters and academic references will be shared with admissions committees on request. Past collaborators include faculty at Jahangirnagar University and the Accident Research Institute (ARI), BUET.

10

Contact

Open to PhD positions, research collaborations, and applied AI work.

Based in Dhaka, Bangladesh (UTC+6). Happy to discuss fit over email or meet on Zoom.

Emailajoy.stu20181@juniv.eduPhone+880-1740-114915LinkedInlinkedin.com/in/ajoysrjuGitHubgithub.com/AJOYSRGoogle Scholarscholar.google.comHugging Facehuggingface.co/ajoysr