Portfolio

Projects & Research

Research systems, papers, and software, from machine learning and autonomous robotics to embedded vision and geospatial analytics.

Comprehensive Evaluation of Machine Learning and Deep Learning Models for Land Use and Land Cover Classification in East Sikkim

East Sikkim, India (~954 km²) Multi-temporal imagery, 2019-2024 Advanced

My flagship AI research project: a publication-style benchmark of 22 machine learning, deep learning, transformer, stacking, ensemble, and feature-fusion approaches for LULC classification over the ecologically sensitive Himalayan district of East Sikkim. The imagery comes from three sources: Sentinel-2 optical, Sentinel-1 SAR, and SRTM topography. Each sample is described by a 24-dimensional feature vector (spectral bands, indices, SAR backscatter, terrain, and texture) extracted via Google Earth Engine. Random Forest achieved the best overall accuracy of 93.44% (Kappa 0.9206, Macro-F1 93.24%).

Models

93.44%

Best OA

0.9206

Kappa

Features

10,530

Samples

LULC Classes

Sentinel-1 SARSentinel-2SRTM Google Earth EnginePyTorchscikit-learn XGBoost / LightGBM / CatBoostTransformersKaggle GPU

Bar chart comparing overall accuracy of all 22 models, with Random Forest highest at 93.44%

S1 + S2 + SRTM GEE Feature Extraction (24-D) 22-Model Zoo 5-Fold CV LULC Maps

Results gallery — click to enlarge

Grouped bar chart of overall accuracy, Kappa, and Macro-F1 for all 22 models

Random Forest confusion matrix across the six LULC classes

Radar chart of per-class F1 scores for top models

Scatter plot of overall accuracy versus Kappa for all models

Fold-by-fold accuracy of the custom Swin-Tiny-Tabular model

Full 22-model leaderboard

Validation results on 3,246 held-out samples; six LULC classes, 24-D feature vector.
#	Model	Category	OA (%)	Kappa	Macro F1 (%)
1	Random Forest	Tree Ensemble	93.44	0.9206	93.24
2	CatBoost	Tree Ensemble	93.01	0.9154	92.84
3	Gradient Tree Boost	Tree Ensemble	92.73	0.9121	92.49
4	XGBoost	Tree Ensemble	92.51	0.9095	92.26
5	LightGBM	Tree Ensemble	92.48	0.9091	92.25
6	Soft Vote (Top-5)	Ensemble	92.30	0.9068	92.15
7	Weighted Soft Vote (Top-5)	Ensemble	92.30	0.9068	92.15
8	1D-CNN	Deep Neural Network	92.17	0.9052	91.90
9	Weighted Soft Vote (Top-3)	Ensemble	92.11	0.9045	91.95
10	MLP	Deep Neural Network	91.90	0.9019	91.77
11	SpectralFormer	Tabular Transformer	91.87	0.9016	91.73
12	Fusion SF+CatBoost (Emb)	Feature Fusion	91.81	0.9009	91.70
13	Fusion All+CatBoost	Feature Fusion	91.77	0.9005	91.65
14	TabTransformer	Tabular Transformer	91.71	0.8998	91.47
15	FT-Transformer	Tabular Transformer	91.53	0.8976	91.41
16	Fusion SF+CatBoost	Feature Fusion	91.53	0.8976	91.43
17	Swin-Tiny-Tabular (custom)	Vision Transformer (1D)	90.83	0.8890	90.75
18	Stacking MLP (Top-5)	Stacking	90.57	0.8862	90.51
19	Stacking LR (Top-5)	Stacking	90.51	0.8854	90.45
20	SVM	Classical ML	90.20	0.8815	90.09
21	KNN	Classical ML	86.72	0.8397	86.64
22	Naive Bayes	Classical ML	84.20	0.8095	84.23

Technical details

Model paradigms: classical ML (Random Forest, SVM, KNN, Naive Bayes), gradient boosting (XGBoost, LightGBM, CatBoost, GTB), deep networks (MLP, 1D-CNN), tabular transformers (TabTransformer, FT-Transformer, SpectralFormer, custom Swin-Tiny-Tabular), and hybrids (soft voting, stacking, feature fusion).
Feature vector (24-D): Sentinel-2 bands B2-B12, spectral indices (NDVI, NDVI-diff, NDWI, MNDWI, NDBI, NDRE, NDSI, NBR, BSI), Sentinel-1 VV/VH backscatter, SRTM elevation/slope/northness, and GLCM texture (contrast, entropy).
Dataset: 10,530 labeled samples (7,284 train / 3,246 validation) across 6 classes: built-up, vegetation/forest, agriculture, water, barren rock, and snow/glaciers.
Custom architecture: Swin-Tiny-Tabular, a 27M-parameter 1D adaptation of the Swin Transformer with windowed self-attention, patch merging, and stochastic depth, trained with 5-fold stratified CV (OOF accuracy 90.83%).
Top results: Random Forest 93.44% OA, CatBoost 93.01%, Gradient Tree Boost 92.73%. Tree ensembles led all paradigms on this tabular spectral task.

TRUST-LLM pipeline: question through Mistral-7B, four uncertainty scores, conformal calibration, trust or flag decision

TRUST-LLM: Conformal Hallucination Detection

Mistral-7B (4-bit) · Kaggle T4×2 2026 Advanced

Can an LLM's uncertainty be turned into a provable guarantee instead of a raw confidence score? I applied split conformal prediction to four uncertainty-scoring methods (log-probability, token entropy, semantic entropy, self-consistency), calibrating a formal trust region on TriviaQA (α = 0.10) and stress-testing it on adversarial TruthfulQA and domain-shifted NQ-Open and MedMCQA — 1,100 questions, 6,600 generations, under three hours on free-tier GPUs with no fine-tuning. Key finding: coverage and detection utility move in opposite directions under shift, and the MedMCQA failure is a concept shift that importance-weighted recalibration provably cannot repair.

0.90

Target Coverage

0.826

Best OOD F1

4 × 4

Methods × Domains

6,600

Generations

Conformal PredictionPyTorch TransformersMistral-7B 4-bit Sentence-Transformers

Results gallery

Bar chart of conformal coverage per method across four test conditions

Bar chart of hallucination detection F1 per method across four test conditions

Heatmap of AUROC for four methods across four QA datasets

GitHub

Overlaid best-trial trajectories comparing manual and autonomous navigation

Communication-Integrated Mobile Environmental Monitoring Framework

Sikkim University & CSIR-CMERI 2025-2026 Advanced

A unified autonomous robotic platform for environmental monitoring: YOLO-seg vision-guided navigation (mask mAP@50 = 0.988) ONNX-optimized to 15.4 FPS on a Raspberry Pi CPU, fused with a tuned PID controller and a full sensor-to-cloud communication stack (Flask, Firebase, ThingSpeak, SQLite).

95.37%

Accuracy

87.4%

RMSE ↓

15.4

FPS on Pi

Trials

Raspberry PiYOLO-segONNX Runtime PID ControlFlaskFirebase SQLiteArduinoESP32

Technical details

PID controller (Kp = 13.2, Kd = 0.475) achieves 95.37% tracking accuracy and an 87.4% RMSE reduction versus manual teleoperation across 40 controlled trials.
YOLO-seg navigation model optimized through ONNX conversion: 15.4 FPS at 64.8 ms latency on Raspberry Pi CPU, a 1.57× speedup.
Communication stack streams multi-sensor environmental data to Firebase and ThingSpeak with SQLite local buffering and a Flask dashboard.

Results gallery

Tracking error over time for the best manual and autonomous trials

Flask dashboard showing live environmental sensor readings

ThingSpeak dashboard with streamed environmental data channels

Waterfall chart showing RMSE improvement from manual to autonomous navigation

Performance comparison of the three machine learning models in the monitoring pipeline

Smart Environment Monitor

Raspberry Pi & Arduino 2026 Intermediate

ML-augmented IoT sensing platform running three models (classification, clustering, and anomaly detection) behind a 6-level alert engine. Data flows from Arduino through a Raspberry Pi into SQLite, Firebase, and ThingSpeak, with a Flask dashboard on top.

scikit-learnFlask FirebaseThingSpeakSQLite

Results gallery

Flask dashboard of the smart environment monitor

ThingSpeak dashboard with sensor data channels

ROC curves for all models in the monitoring pipeline

Feature importance chart of the Random Forest classifier

GitHub Models

In-band accuracy distribution of the hybrid ML and PID controller across runs

Tripathagamini-S Autonomous Robot

CSIR-CMERI, Micro Robotics Lab Feb-Mar 2026 Advanced

A line-following robot built in four versions: v1 YOLO ONNX vision (25-30 FPS), v2 CSV trajectory logging, v3 five-channel IR array at 100 Hz via Arduino, and v4 hybrid ML + PID control reaching 99.6% in-band accuracy. The trained model is published on HuggingFace.

99.6%

Accuracy

97.9%

Error Energy ↓

100 Hz

Control Rate

YOLOscikit-learnPID Control IR SensorsArduinoHuggingFace

Technical details

Hybrid controller: a Random Forest residual-error predictor trained on 208,983 time steps augments a classical PID loop, cutting MAE by 92.1% and error energy by 97.9%.
Lyapunov stability of the hybrid ML + PID architecture formally proven.
Training data (221,967 rows over 100 robot runs) published on Kaggle; Optuna-tuned model (432 MB) on HuggingFace.

Results gallery

Accuracy proof chart showing in-band samples

Chart of the machine learning contribution to control performance

Metrics table comparing PID and hybrid controllers

GitHub Model Dataset

Robot camera frame from the custom path segmentation dataset

YOLO Edge Optimization

CSIR-CMERI 2026 Advanced

Model-to-deployment pipeline for real-time path segmentation on Raspberry Pi. I built a custom 289-image dataset, trained YOLO-seg, and converted it to ONNX at 320×320, reaching 15.4 FPS / 64.8 ms, a 1.57× speedup with zero memory leaks over a 30-minute stability run.

15.4

FPS

64.8 ms

Latency

1.57×

Speedup

YOLO-segONNX Runtime Raspberry PiUltralytics

Results gallery

Benchmark bars comparing PyTorch and ONNX inference time and FPS on Raspberry Pi 5

Three dataset frames with their path segmentation masks overlaid

GitHub

Photo, ground-truth sketch, and U-Net predicted sketch for three test faces

Image to Sketch using U-Net

TensorFlow / Keras 2025 Intermediate

U-Net CNN for photo-to-pencil-sketch transformation trained with a custom MAE + SSIM loss, reaching an 85% SSIM score. Deployed as a Streamlit web app with sub-2-second inference.

85%

SSIM

<2 s

Inference

TensorFlow / KerasU-NetStreamlit

Results gallery

Grid of photos, true sketches, and predicted sketches from the U-Net model

True sketch, predicted sketch, and pixel-wise error heatmap

GitHub

Trajectories of all 40 trials, 20 manual versus 20 autonomous

Manual vs Autonomous Navigation: Experimental Comparison

Sikkim University & CSIR-CMERI Mar-Apr 2026 Advanced

40-trial controlled study comparing autonomous YOLO + PID navigation against human teleoperation: mean RMSE 2.119 cm vs 16.768 cm (87.4% reduction), accuracy 95.37% vs 63.32%, Cohen's d = 8.944.

YOLOHomographyPID Control OpenCVStatistical Analysis

Key statistics (20 + 20 trials)

Metric	Manual	Autonomous	Improvement
Mean RMSE	16.768 cm	2.119 cm	87.4%
Median RMSE	16.211 cm	1.849 cm	88.6%
Std deviation	2.101 cm	0.977 cm	53.5%
Tracking accuracy	63.32%	95.37%	+32.05 pp

Mann-Whitney U = 400, p = 3.40×10⁻⁸; permutation test p < 0.0001; Cohen's d = 8.944 (exceptionally large effect).

Results gallery

Box and violin plot comparing manual and autonomous RMSE distributions

Error over time for best manual and autonomous trials

Summary statistics table of the navigation study

GitHub

RMSE comparison between the baseline PID and the Optuna-tuned hybrid controller

ML-Augmented PID Control

Sikkim University & CSIR-CMERI 2026 Advanced

Hybrid control research: a Random Forest residual predictor trained on 208,983 time steps layered on a PID loop. Results: 92.1% MAE reduction, 97.9% error energy reduction, and 99.6% in-band accuracy, with Lyapunov stability formally proven.

scikit-learnPID Control Lyapunov StabilityArduinoRaspberry Pi

Results gallery

Predicted versus actual scatter plot of the tuned model

Feature importance chart of the tuned model

GitHub Training Repo

Best trial trajectory analysis from homography-based tracking

Tripathagamini-S: Homography Trajectory Tracking

CSIR-CMERI, Micro Robotics Lab 2026 Intermediate

Web-teleoperated robot with homography-based trajectory analysis: 4-point interactive calibration and simultaneous path + robot tracking at 30+ FPS, serving as the ground-truth measurement platform for the navigation study.

HomographyOpenCVFlask PCA9685Raspberry Pi

Results gallery

GitHub

FGSM attack visualization showing clean images, perturbations, and adversarial results

Adversarial Robustness on CIFAR-10

PyTorch 2025 Intermediate

Complete adversarial robustness pipeline: FGSM and PGD attacks implemented from scratch, followed by PGD adversarial training as a defense, with publication-ready robustness visualizations.

PyTorchFGSM / PGDCIFAR-10

Results gallery

Robustness curves under increasing attack strength

Final comparison chart of model robustness

Summary table of adversarial robustness results

GitHub

Confusion matrices of the SVM and MLP scream classifiers

Scream / Non-Scream Sound Detection

Raspberry Pi 2025 Intermediate

Binary audio classification for real-time distress-sound detection, engineered for edge deployment on Raspberry Pi. Trained SVM and MLP classifiers on 26 MFCC-based features extracted from 3,128 audio clips — the SVM reaches 94.89% accuracy (F1 0.95, ROC-AUC 0.986) on the held-out test set.

94.89%

SVM Accuracy

0.986

ROC-AUC

3,128

Audio Clips

PythonMFCC FeaturesSVM / MLPRaspberry Pi

Results gallery

Side-by-side confusion matrices for the SVM and MLP classifiers

ROC curve of the SVM classifier with AUC 0.9861

ROC curve of the MLP classifier with AUC 0.9828

PCA scatter plot of scream and non-scream audio features

GitHub

Classified land cover map of East Sikkim produced in Google Earth Engine

Satellite Land Cover Classification (GEE)

Google Earth Engine 2025 Intermediate

Random Forest land cover mapping over East Sikkim with multi-year Sentinel-2 composites: 92% accuracy across five classes over 500+ km², with an automated GEE + QGIS workflow.

Google Earth EngineSentinel-2 NDVI / NDWI / NDBIQGIS

Results gallery

Accuracy summary table comparing classifiers

GitHub

NDWI water and non-water classification map of Chilika Lake

Water Body Mapping: NDWI & Sentinel-2

Google Earth Engine 2025 Intermediate

Automated water body detection using NDWI on Sentinel-2 imagery over Chilika Lake, Odisha. The GEE + QGIS workflow delineated 3,264 individual water bodies totalling ~55,823 hectares from the 2023 composite, with full raster-to-vector statistics on the size distribution.

3,264

Water Bodies

55,823 ha

Mapped Area

NDWIGoogle Earth Engine Sentinel-2QGIS

Results gallery

Binary water and non-water raster map of Chilika Lake

Histogram of water body sizes with summary statistics

Study area map of Chilika Lake with India inset

GitHub

CSV Sensor Logger

Raspberry Pi 2025 Beginner

Flask-based sensor monitoring utility for Raspberry Pi with live readings and one-click CSV export. This is the data-capture backbone reused across my robotics experiments.

FlaskPySerialCSV

GitHub