HomeEducationSkills ProjectsExperienceCertifications CVContact

Projects

Research systems, papers, and software spanning autonomous robotics, embedded AI, machine learning, and geospatial analysis.

Research Paper

Manual vs Autonomous Robot Navigation: Experimental Comparison

Sikkim University & CSIR-CMERI  ·  March - April 2026

40-trial study: autonomous YOLO+PID vs human teleoperation. Results: mean RMSE 2.119 cm vs 16.768 cm, 87.4% reduction. Accuracy 95.37% vs 63.32%. Cohen's d = 8.944.

YOLOHomographyPID ControlOpenCVStatistical AnalysisRaspberry Pi
Hybrid PID

ML-Augmented PID Control for Autonomous Line-Following Robot

Sikkim University & CSIR-CMERI  ·  2026

Hybrid control: Random Forest trained on 208,983 time steps. Results: 92.1% MAE reduction, 97.9% error energy reduction, 99.6% in-band accuracy. Lyapunov stability formally proven.

scikit-learnPID ControlLyapunov StabilityArduinoRaspberry Pi
4-Version Evolution

Tripathagamini-S Auto / Autonomous Line-Following Robot

CSIR-CMERI, Micro Robotics Lab  ·  Feb - 2026

v1.0.0: YOLO ONNX at 25-30 FPS. v2.0.0: CSV trajectory logging. v3.0.0: 5-channel IR array via Arduino at 100 Hz. v4.0.0: Hybrid ML+PID, 99.6% accuracy.

YOLOscikit-learnPID ControlIR SensorsArduinoHuggingFace
Teleoperation + Vision

Tripathagamini-S / Homography Trajectory Tracking Platform

CSIR-CMERI, Micro Robotics Lab  ·  2026

Web-based teleoperated robot with homography-based trajectory analysis. 4-point interactive calibration, simultaneous path and robot tracking at 30+ FPS. Ground-truth measurement platform.

HomographyOpenCVFlaskPCA9685Raspberry Pi
Embedded Optimization

YOLO Path Segmentation & Real-Time Optimization

CSIR-CMERI  ·  2026

Complete model-to-deployment pipeline. 289-image custom dataset. 320x320 ONNX: 15.4 FPS, 64.8 ms (1.57x speedup). 30-min stability: 15.2±0.8 FPS, zero memory leaks.

YOLO-segONNX RuntimeRaspberry PiUltralytics
ML Training Repo

pid-ml-follower / RF Residual Error Predictor

CSIR-CMERI  ·  March 2026

221,967 rows, 100 robot runs. Optuna tuning (50 trials). Tuned RF: RMSE = 0.018731, 14.3% over persist-error baseline. Model (432 MB) on HuggingFace.

scikit-learnOptunaKaggleHuggingFace
IoT + ML Pipeline

Smart Environment Monitor / ML-Augmented IoT Sensing

Raspberry Pi & Arduino  ·  2026

Three ML models: classifier, clustering, anomaly detection. 6-level alert engine. Data: Arduino → RPi → SQLite + Firebase + ThingSpeak + Flask.

scikit-learnFlaskFirebaseThingSpeak
Adversarial ML

Adversarial Robustness on CIFAR-10 / FGSM & PGD

PyTorch  ·  2025

Complete adversarial robustness pipeline on CIFAR-10. Implements FGSM and PGD attacks, then PGD adversarial training as defense. Publication-ready visualizations.

PyTorchCIFAR-10
Deep Learning

Image-to-Sketch Generation using U-Net CNN

TensorFlow  ·  2025

U-Net CNN for image-to-sketch transformation with custom MAE+SSIM loss achieving 85% SSIM score. Deployed as Streamlit web app with under 2-second inference.

TensorFlow / KerasStreamlit
Audio ML

Sound Detection / Scream & Non-Scream Classification

Raspberry Pi  ·  2025

Binary audio classification system for real-time distress sound detection targeting edge deployment on Raspberry Pi.

PyTorchRaspberry PiPython
Remote Sensing

Satellite Land Use / Land Cover Classification

Google Earth Engine  ·  2025

ML classifier for multi-class land cover mapping. 92% accuracy across five land classes over 500+ km². Automated GEE workflow.

Google Earth EngineSentinel-2NDVI/NDWI/NDBIQGIS
Remote Sensing

Water Body Mapping using NDWI & Sentinel-2

Google Earth Engine  ·  2025

Automated water body detection using NDWI on Sentinel-2. Multi-temporal analysis tracking seasonal water extent variations.

NDWIGoogle Earth EngineSentinel-2QGIS