Spatio-temporal Kriging & Forecasting (Thesis)

• Designed spatio-temporal PM2.5 forecasting & interpolation pipelines for sensor networks across Bihar and Uttar Pradesh, using ST-LLM, Geostatistical Kriging, Tree based methods (XGBoost, etc.) and diffusion-driven Inductive GNNs.
• Developed and deployed production-ready APIs powering a real-time pollution maps & hotspot detection system at 500m x 500m grid, to enable early-warning alerts and forecast-assisted policy interventions.
• Currently extending the research by benchmarking additional foundational models to improve generalizability. Latest experiments demonstrate RMSE less than 40 μg/m3 and MAPE ≈ 30% on real sensor data, indicating strong predictive reliability.

Forecasting using ST-LLM+ Graph Neural Networks + PEFT-LoRA Finetuning on GPT2

• Hybrid ST-LLM + GNN Framework: Designed a novel pipeline where a Spatio-Temporal LLM captures long-range temporal dependencies, while a Graph Neural Network (GNN) encodes spatial correlations across sensors. This joint modeling significantly improved PM2.5 forecasting accuracy compared to standalone temporal or spatial models.
• Parameter-Efficient Fine-Tuning (PEFT): Applied LoRA-based PEFT on ST-LLM layers instead of full fine-tuning, reducing trainable parameters by >90% while preserving forecasting performance. This enabled efficient adaptation to domain-specific air quality data on limited GPU resources.

Data Imputation with Generative AI- Conditional GANs

• Built a cGAN-based imputation framework where the generator reconstructs missing PM2.5, Temp, RH values conditioned on spatio-temporal features (latitude, longitude, time encodings, and neighbor sensor averages). This was used in preparing imputed data to used in different ML models and problems later
• Time Series Clustering and Airshed

• Built an LSTM-based temporal embedding model for PM2.5 time series, extracting latent pollution dynamics from multiple sensor locations to represent long-term spatio-temporal dependencies.
• Applied hierarchical agglomerative clustering (Ward’s linkage + dendrogram cuttings) on LSTM embeddings to identify natural airsheds, grouping locations with shared pollution signatures beyond political boundaries.
• Developed interactive Plotly dashboards and dendrogram visualizations to explore seasonal airsheds, enabling policymakers to interpret clusters and design regionally coordinated clean-air strategies.

Agent-Based Intelligent Tutor Chatbot

• Built a conversational AI chatbot with LLaMA 3.2 via Ollama and Langchain.
• It was backed by RAG pipeline (Real-Time web search and uploaded PDFs via FAISS indexing), serving student’s academic queries with a personalized chatbot.
• Established modular agent-based system to get multi-hop reasoning and tool-use like (Web search, file reader).

Doodle-Vision: Air-Drawn Sketch Recognition

• Real-time sketch classification on QuickDraw dataset with hybrid ConvLSTM, Vision Transformers, and CNN-Transformer architectures for multimodal learning. Achieved 85% accuracy on 50+ classes.
• Created a real-time drawing interface to capture hand drawn sketches using webcam for object recognition; also explored lightweight models: MobileNet for achieving high accuracy suitable for mobile deployment.
• The project work is also in pipeline to be submitted for publication in WACV 2025 Conference.

Tuning LLM via PEFT for IMDB Reviews

• Fine-tuned T5-small with PEFT methods (Prompt Tuning, LoRA) for sentiment classification.
• Reduced trainable parameters by ~80% with minimal accuracy loss, making it suitable for edge inference.
Tech stack: PEFT, LoRA, HuggingFace, Transformers, Adapters
View Project

Topic Mining on IRCTC Reviews

• Scraped Google Playstore Reviews for IRCTC App using Selenium and Beautiful Soup libraries.
Implemented topic mining on it using LDA, BERT embeddings, UMAP dimensionality reduction and HDBSCAN clustering for identifying key user concerns and improving service; Identified 66+ niche topics and visualized using t-SNE plots.

AmazoLens: Amazon Sales Analysis

• Built recommendation systems, market basket analysis (Apriori, FP-Growth), and customer segmentation from Amazon sales dataset.
• Created interactive dashboards with Plotly/D3 and ClickHouse backend.

CRM Scoring Improvement Classplus, Noida

• Designed predictive scoring model for telesales at Classplus/Testbook.
• Increased sales conversion by 15%+ by identifying high-intent student leads.

Project Hedwig: Text Classification Classplus, Noida

• Deployed model to classify 200+ categories of educational content with 94% accuracy, reducing manual effort and saving KYC costs.

Exam Seasonality Analysis Classplus, Noida

• Built demand forecasting pipeline for exam-specific materials. Triggered sales alerts from Google API + transactional data to optimize strategies.