Transforming Research into Real-World Impact

A reproducible, explainable transformer pipeline for depression emotion/severity experiments, including ablations, XAI faithfulness checks, and a minimal Flask demo (research use only).

Lightweight hybrid CNN–Transformer (MobileViT + attention + texture cues) for efficient and explainable lung cancer diagnosis on CT/histopathology with Grad-CAM and robust evaluation support.

A modular pipeline for audio-visual object recognition using hybrid, tensor, and FiLM-style fusion with flexible feature extraction and noise-robust training options.

Flask-based web application for cotton leaf disease, fabric stain defect detection, and fabric composition classification with probability charts and Grad-CAM explanations.

MaxViT-based soybean leaf/seed disease classification web app with Grad-CAM heatmaps, probability visualization, and a clean UI for practical screening workflows.

A risk-aware clinical decision support pipeline that fuses medical images with radiology notes and structured signals to produce calibrated predictions, uncertainty flags, and evidence-grounded outputs for safer triage and reporting assistance.

A standardized evaluation suite to measure and improve model performance across hospitals, scanners, and patient subgroups—supporting domain-shift testing, fairness slices, and reproducible reporting for deployment-ready medical AI.

A safety layer that detects uncertain, out-of-distribution, or artifact-corrupted cases and defers them for human review. Includes coverage–risk analysis, abstention policies, and audit-friendly logs for high-stakes clinical deployment.

A clinician-oriented explainability toolkit that goes beyond heatmaps—providing concept-based explanations, counterfactual evidence, faithfulness checks, and concise explanation report cards to support transparent and auditable medical AI.

A safety-first LLM workflow that drafts structured clinical summaries using only verified evidence (model outputs, metadata, and approved templates). Includes confidence-aware refusal, traceable citations, and guardrails for responsible use.

A privacy-preserving foundation model trained across institutions without centralizing patient data. Focuses on federated optimization, calibration under client shift, and robust performance across sites and scanners.

Risk forecasting from serial scans to predict progression and time-to-event outcomes (e.g., glaucoma progression). Produces calibrated risk curves, uncertainty, and clinician-friendly timelines for follow-up planning.

A quality-control layer that detects motion blur, low contrast, compression, and device artifacts before inference. Routes low-quality cases for re-capture or robust enhancement to reduce silent failures in practice.

A monitoring and evaluation dashboard that reports performance, calibration, and failure modes across age/sex/site/device subgroups. Includes bias discovery, shift alerts, and standardized reporting for responsible deployment.

A self-supervised learning pipeline (contrastive/masked modeling) for data-efficient medical imaging. Targets strong transfer across modalities with minimal labels and robust generalization under dataset shift.

Counterfactual and concept-based explanations designed for clinical reasoning—testing what minimal, plausible image changes would alter predictions while tracking faithfulness and safety constraints.

A lightweight, real-time ultrasound screening pipeline optimized for low-resource clinics. Focuses on efficient architectures, quantization, and a simple operator-facing interface with uncertainty-aware alerts.