Projects
My projects in artificial intelligence, language models, computer vision and electronics.
Certified Segmentation Update Selection
Developed an exact mask-pair Dice envelope and a budgeted return policy with logit interpolation. The manuscript reports validation against 5,026,592 reference masks and an evaluation of six trained sources across 17,808 case-condition-seed episodes, distinguishing admissibility from adaptation benefit.
Retinal VLM Class Balancing & External Transfer
Audited the archived IDRiD-VQA pipeline, cleaned the local image split, and compared two sampling strategies at three paired seeds. Evaluated all six final checkpoints on 83 IDRiD validation images and 400 external RetinaMNIST+ images, reporting per-grade errors, uncertainty, and the trade-off between balanced and overall accuracy.
Actual retinal images; scanning guides are illustrative.
IDRiD-VQA: Medical Visual Question Answering
Fine-tuned Qwen2-VL-2B on IDRiD with synthetic, label-derived question-answer targets and built a Gradio interface. The October follow-up audits this archived pipeline and evaluates a corrected grade-only adaptation; generated explanations do not establish lesion grounding.
Illustrative grading workflow, not a recorded scoring result.
Automated Subjective Answer Sheet Evaluation
A multimodal system for handwritten-text extraction, semantic scoring, concept-based grading, diagram analysis, and examiner review. The project reduced manual evaluation time by over 70%. Supervised by Dr. Ashwini Kumar Upadhyay, REC Kannauj.
Saved-model demonstration; predictions are not fact checks.
Actual detections; crash and notification are a separate simulation. Footage: Lindsey Mateo / Pexels.
Conceptual architecture illustration.
Model Efficiency with Hybrid RNN–Transformer Architectures
Exploring hybrid recurrent and attention-based architectures for language modeling in resource-constrained environments, targeting parameter count, memory footprint, and inference latency.
Illustrative workflow, not a live agent run.
AI Tool for Data and Web Insights
An interface for uploading datasets, creating queries, running web searches, and using language models to extract and synthesize information. Results can be reviewed and downloaded as structured data.
Illustrative text classification, not a clinical diagnosis.
Multi-Label Mental Health Text Classification
A BERT-based multi-label model for identifying anxiety- and depression-related categories in text, using contextual embeddings and attention mechanisms.
New YuNet/SFace demonstration, not the original FaceNet implementation. Photo: NASA via scikit-image.
Face Recognition using Deep Learning
A face-recognition system using OpenCV and FaceNet embeddings for identity verification on trained facial data.