Arnesh Banerjee

Advisor: Dr. Sourangshu Bhattacharya

• Working on IgLM, India's first population-specific genomic foundational model (StripedHyena2 architecture), under the GRISHMA Summer Internship Program. Reached pooled XGBoost ROC-AUC 0.933 on single-sample oral-cavity cancer detection from RNA-seq (58,147 genes, 1,206 samples across 1,160 patients) by training an L1-embedded classifier family under patient-grouped 5-fold stratified cross-validation with predictions pooled out-of-fold.
• Verified the signal is biological — not tissue-of-origin — by designing a layered TCGA negative class (non-oral HNSC, solid-tissue normal, 8 unrelated cancers), reporting per-stratum sub-AUCs of 0.800, 0.900, and 0.987; confirmed robustness to class skew via a 1.5:1 balanced re-run preserving the headline (XGBoost 0.928).
• Ported the Cedars-Sinai Molecular-Twin (MTPilot) L1-embedded model family into a shared module reused across the IgLM detection and downstream survival-prediction pipelines.

Advisor: Prof. Debotosh Bhattacharjee

• Designed a hybrid CNN–Transformer segmentation model (ResNet-34 encoder + ASPP + Transformer bottleneck + SE-gated skip connections) with a differentiable Chan–Vese level-set loss, reaching 0.9716 Dice and 0.9463 IoU on the DMR-IR dataset (357 thermograms, 119 patients) under patient-stratified 5-fold cross-validation.
• Surfaced an annotation-quality ceiling in weakly supervised thermography by benchmarking against four SOTA baselines (Attention U-Net, UNet++, DeepLabV3+, TransUNet) on five metrics (Dice, IoU, HD95, ASSD, BF1) and showing all models converge to statistically indistinguishable Dice (≈0.97, p > 0.05, paired Wilcoxon with 1000-resample bootstrap CIs).
• Built a robustness battery (label-noise injection at 10–30%, augmentation regimes, 25–100% training subsets) and an explainability suite (Grad-CAM, attention maps, Monte-Carlo dropout uncertainty) for clinician-facing decision support.

Advisor: Dr. Arnob Ghosh

• Built a 2,500-pair safe/unsafe prompt–response dataset spanning jailbreak strategies, indirect requests, role-play, multi-step instructions, and ethical/unethical educational queries; assigned absolute binary harm labels (replacing the Bradley–Terry pairwise scheme) and fine-tuned the final six layers of LLaMA-2-7B-chat-hf with a dense classification head as the CS-RLHF cost model.
• Validated semantic grounding of the cost model on the held-out test split and the external XS-Test benchmark, reaching ≈92% alignment with human safety judgments and XS-Test scores of 0.91–0.96 (matching human verdict 0.89–0.92), versus 0.07–0.32 for the Safe-RLHF baseline cost model.
• Demonstrated the trained policy is 8× more efficient at flagging unsafe responses than Safe-RLHF and is preferred by humans in ≈60% of head-to-head comparisons (+70 Elo) over 1,000 sampled prompts; co-authored the resulting COLM 2026 submission (arXiv:2510.03520).

Advisor: Ms. Arpita Talukdar

• Improved WPBC accuracy to 93.67% (SVM + RFE) and WDBC to 97.77% (LogReg + RFE) by adding RFE/SFS feature selection, SMOTE class balancing, and GridSearchCV hyperparameter tuning over a dual-stage diagnosis-and- recurrence ML framework benchmarking five classifiers (RF, SVM, Logistic Regression, MLP, XGBoost) under stratified 10-fold cross-validation.
• Identified clinically relevant nuclear features through a comparative analysis across model–feature-selection combinations, reported with bootstrap confidence intervals.