Ajoy Sarker

Problem

Emotion from EEG is a robustness problem: non-stationary signals, artifacts, subject/session shift, and long-range temporal structure across multichannel data.

Approach

• Preprocessing — bandpass filtering, ICA, artifact removal
• Band-power features across delta–gamma; CNN / RNN / Transformer baselines
• Cross-subject and cross-session evaluation on DEAP and MAHNOB-HCI

Finding

Feature-rich representations outperform linear baselines on overlapping EEG classes; kernel SVMs and wide ANNs reach up to ~98.4% test accuracy in reported runs, while simpler models break down under class overlap.

Problem

Reliable moderation for a low-resource language: severe class imbalance, rich morphology, and code-switching—where off-the-shelf English models fail.

Approach

• Multi-level TF–IDF (word + character n-grams)
• Stacked ensemble: LinearSVC, logistic regression, XGBoost
• Extended BD-SHS to a 15-class taxonomy (~58k samples)

Finding

Stacked ensemble reaches 91.2% accuracy and 90.7% macro-F1 on 15-class Bengali social media—about +1.9 points over prior work.

Problem

Production LLM systems fail silently—hallucinated answers, brittle retrieval, and unvalidated tool use. Deployment needs measurable reliability, not demos.

Approach

• RAG with dense + sparse retrieval (pgvector) and latency/accuracy evaluation
• LangGraph workflows with structured outputs and output guardrails
• Schema-constrained validation on multimodal KYC/document pipelines

Finding

92% retrieval accuracy at sub-200 ms latency in production RAG; 1st place at company AI hackathon for guardrailed hybrid search.

Problem

Tabular ML under noisy, imbalanced administrative records—predicting crash occurrence and severity across all 64 districts of Bangladesh.

Approach

• Gradient-boosted trees: CatBoost, XGBoost, Random Forest
• Feature engineering across demographic, environmental, and infrastructure fields
• Class imbalance via SMOTE and cost-sensitive learning

Finding

CatBoost leads occurrence prediction (F1 ≈ 0.65) on one of Bangladesh's largest longitudinal crash datasets (500K+ records, 2001–2020).