Undergraduate Researcher · B.Tech CSE (Data Science)
Heritage Institute of Technology, Kolkata

• Selected for the GRISHMA Summer Internship Program at IIT Kharagpur; working on India’s first genomic foundational model (IgLM) — a population‑specific genomic language model built on the StripedHyena2 architecture.

• Achieved 98.10% Dice and 96.39% IoU on breast thermography tumor segmentation using a ResNet34 + ASPP + ViT decoder with SE attention, evaluated via 5‑fold cross‑validation.
• Conducted a rigorous 20‑run ablation study across U‑Net, ResNet‑only, and hybrid architectures, revealing generalization trade‑offs in medical imaging.
• Built an automated cross‑validation pipeline with checkpointing, experiment logging, and reproducible training for publication‑ready metrics.

• Constructed a cost‑model training dataset for a modified Safe RLHF pipeline, leading to improved cost‑model performance.
• Designed and tested a comparative 800‑datapoint safety benchmark against state‑of‑the‑art LLMs (GPT, DeepSeek, Gemini, Qwen, Mistral).
• Demonstrated 50% higher safety compliance than ChatGPT 5 across benchmark scenarios.

• Developed a dual‑stage ML framework for breast cancer diagnosis and recurrence prediction using WDBC and WPBC datasets, evaluating RF, SVM, Logistic Regression, MLP, and XGBoost with stratified 10‑fold cross‑validation.
• Implemented RFE and SFS feature selection with SMOTE and GridSearchCV tuning, achieving 93.67% (WPBC) and 97.77% (WDBC) accuracy.
• Conducted comparative performance analysis and interpreted clinically relevant nuclear features with confidence intervals.