Research, shipped as working systems.

A stacking ensemble predicting crop yields across India using 24 years of agronomic data and NASA POWER climate data, with a phenology-segmented explainability framework and a full-stack farmer advisory dashboard. Accepted for publication.

LLM buyer/seller negotiation agents (Claude Sonnet 4.5) paired with a symbolic layer of constraint checking and game-theoretic solving (Nash equilibrium, Pareto optimality), with a full audit trail and explainability engine.

Middleware that routes ML workloads between real QPU and classical simulator backends — classifying jobs by circuit depth and noise tolerance, scheduling with async dispatch via Ray, and aggregating results into one unified tensor.

A production-level agricultural advisory system combining FastAPI, ChromaDB, and Gemini Flash in an end-to-end RAG pipeline, grounding crop-specific advice in curated agronomic sources — plus a ResNet-based multimodal disease classifier. Accepted at IEEE ICoICET-2026.

Implementation of causal counterfactual explainability methods for tabular ML models, benchmarked across multiple baselines — published at IEEE IC3SIS-2026.

Apology sincerity detection using a hybrid BERT model, supporting multilingual text (English, Hindi, Hinglish) — published at IEEE ICETICS-2026.

Full-stack healthcare AI agent for postpartum depression (PPD) risk prediction, combining RAG-based clinical Q&A with a predictive risk model.

Cooperative multi-agent reinforcement learning system for cloud resource scheduling, outperforming standard baseline schedulers.

Multi-agent reinforcement learning and federated/privacy-preserving ML for industrial circular economy optimization, with Dr. Madhvi Saxena and collaborators from Vishwakarma University and Chandigarh University. Accepted at IDICAIHEI-2026 (November 27–28, 2026).

A modular, swappable-backend RAG pipeline — loading, chunking, embedding, storage, and retrieval as independent pluggable stages, so switching FAISS for Chroma is a one-line change.

A single, consistent API that wraps SHAP, LIME, and Captum behind one interface — explaining or comparing all three takes one line of code instead of three code paths.

A CLI tool that quantifies how well SHAP and LIME agree on the same trained model, using rank correlation and overlap diagnostics.

A lightweight local tool for tracking academic paper submissions across conferences and journals, with one-command export to Markdown or CSV.

Keeps a personal site's publications page and BibTeX file in sync with an ORCID record automatically, runnable on a schedule via GitHub Actions.