Nishtha
Soni.

Citizens have no efficient way to report civic issues to municipalities, leading to slow resolution, duplicate reports, and zero accountability.

React PWA → Node / Express REST API → MongoDB Atlas with geospatial indexing → Role-based middleware → RESTful API layer with optimised query caching.

• AI-powered autofill and structured issue-submission flows, cutting complaint-reporting effort by ~60–70% and manual data-entry fields by ~50%
• Automated civic-issue categorization via NLP/ML at ~85–90%+ classification accuracy
• Real-time tracking of 1,000+ complaints via map-based visualisation and geolocation mapping across 50+ locations
• Role-based access control: citizen, field officer, administrator
• Dashboards monitoring 5+ key metrics (volume, category distribution, resolution status, priority, geographic hotspots) plus community upvoting for citizen-driven prioritisation

Presented to the Smart India Hackathon 2025 finals judging panel, which reported no further improvement recommendations — finishing 1st Runner-Up among 2L+ national submissions.

Scaling geospatial queries to 1,000+ concurrent reports required MongoDB index tuning and query caching, yielding a 70% performance gain.

Law enforcement manually reviews hours of CCTV footage to locate missing individuals — slow, error-prone, and resource-intensive.

Video ingestion → Frame extraction (OpenCV) → Face detection → Embedding comparison (face_recognition / dlib) → Timestamp logging → Alert output.

• 94% identification accuracy across varied lighting and occlusion
• Automated frame-wise detection with precise timestamp logging
• Reduced manual effort by 70%; hours processed in minutes
• Multi-camera and live stream extensibility

Consistent accuracy under poor CCTV lighting required histogram equalisation and CLAHE preprocessing before embedding extraction.

How can Transformer-based NLP decode the emotional complexity of Early Modern English, mapping how Shakespearean characters evolve across acts and scenes?

Domain-adapted corpus → Fine-tuned DistilRoBERTa & DistilBERT → Six-class emotion classifier & sentiment model → Scene-level aggregation → Arc visualisation pipeline.

• Fine-tuned DistilRoBERTa and DistilBERT on Early Modern English for domain adaptation
• Six-class emotion classification at 88.1% accuracy and sentiment analysis at 74.3% accuracy
• Temporal arc mapping across all acts and scenes
• Character-level and play-level emotion heatmaps

Early Modern English deviates significantly from modern corpora. Targeted fine-tuning on digitised Shakespearean texts substantially improved token-level emotion attribution.

Reliable MRI brain-tumor detection demands both high accuracy and clinical interpretability, yet limited medical datasets and class imbalance make robust, trustworthy models hard to train.

MRI input → EfficientNet-B3 backbone → custom classification head → Focal Loss with label smoothing → Grad-CAM++ localisation → interpretable heatmap output.

• Binary MRI brain-tumor classifier on EfficientNet-B3 with a custom classification head, reaching 98.03% test accuracy and AUC 0.9994 on a 305-image held-out test set
• Focal Loss with label smoothing, MixUp augmentation, and CosineAnnealingWarmRestarts to handle class imbalance and improve generalisation on a limited medical dataset
• Grad-CAM++ for pixel-level tumor localisation, producing clinically interpretable heatmaps to support radiologist trust in AI predictions
• Zero false positives, a 2.84% false-negative rate across 305 held-out test images, Cohen’s Kappa of 0.97, and F2 score of 0.99

Outperformed ResNet50, DenseNet201, and Inception baselines while delivering explainable, clinically actionable predictions.

Cross-camera person re-identification must match individuals across viewpoints and lighting, where rigid manual thresholds on deep features generalise poorly.

Person image → part-based split (head, torso, legs) → ResNet-50 embeddings + HSV histograms + HOG descriptors → weighted feature fusion → GMM-calibrated Type-1 Mamdani fuzzy inference → ReID decision.

• Type-1 Mamdani fuzzy inference system integrated with deep-learning features for cross-camera surveillance ReID
• Part-based pipeline dividing person images into head, torso, and leg regions; ResNet-50 embeddings, HSV color histograms, and HOG descriptors
• Weighted feature-fusion framework with fuzzy membership functions calibrated via Gaussian Mixture Models (GMM) to eliminate manual threshold tuning
• SMOTE-based class balancing and decision thresholds optimised with a combined Youden-J and Maximum-F1 strategy

Evaluated on the PETS 2009 benchmark: 99.25% Accuracy, 99.27% ROC-AUC, 98.75% mAP, and 100% Rank-1 Accuracy.