Darshan Joshi — Research Publications

Abstract

This paper presents a production-grade classroom behavior detection framework trained on the SCB-05 dataset using YOLOv8x on Google Colab Pro (A100 GPU). The system achieves 74.85% mAP@0.5 overall across 11+ behavioral classes, with a standout 93.5% detection accuracy on sleeping behavior. A tiled inference pipeline enables real-time deployment on CCTV/RTSP feeds. Grad-CAM explainability layers validate that the model attends to posture and body position rather than facial features, supporting responsible AI deployment in educational settings.

Key Contributions

• ›Trained YOLOv8x on the SCB-05 dataset using Google Colab Pro (A100 GPU), achieving 74.85% mAP@0.5 across 11+ behavioral classes.
• ›Achieved 93.5% detection accuracy on sleeping behavior — the highest per-class precision in the model.
• ›Built tiled inference pipeline for real-time CCTV/RTSP feed processing for live classroom monitoring.
• ›Curated and annotated the SCB-05 multi-class dataset via Roboflow with rigorous quality control.
• ›Integrated Grad-CAM heatmap visualization confirming model attends to posture over facial features.
• ›Proposed RTSP-based deployment architecture for resource-constrained campus hardware.

Keywords

Abstract

This paper presents a hybrid CNN-LSTM deep learning architecture for earthquake prediction and the generation of synthetic seismograms. The model leverages convolutional layers for spatial feature extraction from seismic waveform data and LSTM layers for temporal sequence modeling, enabling accurate magnitude prediction and realistic synthetic seismogram synthesis for data augmentation and simulation purposes.

Key Contributions

• ›Designed a hybrid CNN-LSTM architecture that combines convolutional spatial features with LSTM temporal sequence modeling for seismic data analysis.
• ›Developed a pipeline for processing real seismic waveform datasets and training multi-target prediction models for magnitude and location estimation.
• ›Implemented synthetic seismogram generation capabilities to augment training datasets for improved model generalization.
• ›Evaluated model performance against baseline deep learning models on established seismic datasets.
• ›Demonstrated that CNN-LSTM hybrid models outperform standalone CNN or LSTM architectures on temporal seismic prediction tasks.
• ›Published findings in the American Journal of Civil Engineering, October 30, 2025.

Keywords

Abstract

This paper describes the architecture, design decisions, and evaluation of a production-grade AI job application system built with FastAPI. The system features dual extraction (NLP + LLM), ATS scoring and sanitization, adaptive learning via term memory, recruiter outreach automation, and a full web dashboard. Achieves 87%+ suitability scoring accuracy and sub-2-second generation latency across PDF, DOCX, and JSON outputs.

Key Contributions

• ›Architected a 25+ module production system with FastAPI backend, web dashboard, SQLite database, and full test suite.
• ›Implemented dual JD extraction pipeline: regex/NLP heuristics (jd_extract.py) + LLM-powered extraction (llm_extract.py).
• ›Designed ATS scoring engine (ats_check.py, ats_sanitize.py) ensuring 91/100 ATS compatibility on generated resumes.
• ›Built adaptive learning system (learner.py, term_memory.py) that improves skill matching accuracy over repeated use.
• ›Automated recruiter outreach with personalized email generation (recruiter_msg.py) from JD context.
• ›Benchmarked 87%+ suitability scoring accuracy and sub-2-second generation latency across PDF, DOCX, and JSON.

Keywords