Research & Applied Projects

Developing a hybrid RNN–Transformer model for low-resource language modeling, optimizing parameter use, memory footprint, and inference latency without sacrificing accuracy.

Fine-tuned Qwen2-VL-2B with 4-bit QLoRA (Unsloth) on the IDRiD dataset to grade diabetic retinopathy (0–4) and generate structured clinical explanations — trained end-to-end on a single 6 GB consumer GPU in ~90 minutes. Lifts grading accuracy from 31.1% to 66.0% and QWK from 0.00 to 0.76 (substantial agreement), with a 2,064-pair clinically grounded VQA dataset, offline Gradio demo, and an ACL-format paper. Model and dataset released on Hugging Face.

End-to-end fake-news text classifier: cleaning, TF-IDF features, and a model bake-off that selects a Passive-Aggressive classifier reaching 0.99 F1 / 0.9995 AUC on a held-out test set. Fully reproducible (single seed, deterministic solvers), unit-tested with GitHub Actions CI, a ready-to-deploy Gradio demo, and an honest write-up on source-leakage and why the model learns writing style — not facts.

AI-powered grading system using BERT and SBERT for semantic scoring of descriptive answers. Integrates OCR for handwritten text extraction and reduces manual evaluation time by over 70%. Built under Dr. Ashwini Kumar Upadhyay (REC Kannauj).

AI-driven system to detect road hazards, lane deviations, and obstacles for biker safety using deep learning and computer vision.

BERT-based multi-label classifier that detects mental health disorders like anxiety and depression from user posts using attention mechanisms.

LLM-powered tool that processes datasets, performs intelligent web searches, and extracts user-defined information through a clean interface.

Safety system using YOLOv8 for vehicle detection, speed estimation, and collision risk prediction with real-time alerts.

CNN-based face recognition model using OpenCV and FaceNet embeddings for accurate identity verification.