DEXTER is a local-first, agentic application security assessment CLI built around a real Scan → Analyze → Verify → Refine loop instead of single-pass scanning. A ReAct tool-calling agent investigates ambiguous findings before trusting them — reading code, checking secret entropy, and grepping the repo — across 14 integrated tools spanning SAST, secret scanning, dependency CVEs, and authorization-gated recon/DAST, with every tool run sandboxed in a disposable Docker container.

Credé is an explainable AI credit-scoring platform with separate portals for borrowers and banks. An XGBoost model predicts creditworthiness from user and financial profile data, with SHAP-based explanations breaking down exactly which factors drove each score. Features include OCR-based document ingestion, a bank-side applicant dashboard with fraud alerts, loan application workflows, and a live risk engine — all served through a Flask API and a React dashboard.

MailJi is an AI-powered Gmail spam classifier that ships as both a React + FastAPI web app and a Chrome Extension injecting a live classifier into Gmail's sidebar. A TF-IDF + Multinomial Naive Bayes model (99.3% accuracy on the Enron dataset) flags spam, while LIME explainability highlights the exact words that triggered each verdict. Includes Google OAuth login, a donut-chart spam/ham breakdown, paginated email browsing, and a 5-minute result cache in the extension for fast, low-cost re-scans.

CardioXAI a clinically interpretable machine learning system for heart disease prediction designed for clinical decision support. The system uses Logistic Regression for prediction and SHAP (SHapley Additive exPlanations) to generate interpretable explanations for each case.The primary objective is to balance predictive performance with explanation stability and clinical interpretability, providing healthcare professionals with transparent AI assistance they can trust.

Satellite-Based Crop Recommendation System for Western Maharashtra — integrates Sentinel-2 NDVI imagery and SoilGrids APIs, achieving 92.7% crop recommendation accuracy across 6+ districts. Soily is a powerful digital soil mapping platform that leverages satellite imagery to analyze and predict soil properties such as pH, nutrients, and moisture content. Based on these insights, it provides smart crop recommendations along with the ideal fertilizer suggestions to enhance productivity and maintain soil health.

Serverless P2P File Transfer & Messaging for Local Networks — zero-cloud WebRTC data channels with mDNS/Zeroconf auto-discovery, AES-GCM 256-bit encryption, and FastAPI WebSocket signaling. LanShare is a fully local, serverless application that lets devices on the same Local Area Network (LAN) discover each other automatically, transfer files, and chat in real time — entirely without internet access or a central server. Each device runs its own instance of LanShare. Devices find each other via mDNS (Zeroconf), exchange files directly

QuikKirana is a full-stack kirana (grocery) store management platform covering online storefront, cart, and checkout alongside a back-office suite for inventory, low-stock alerts, billing, and marketing. Built on a Node.js/Express/MongoDB backend with models for products, orders, bills, customers, and shop locations, it gives small store owners a single system to run both their online store and day-to-day shop operations.

LeetCo is a Chrome extension that surfaces company-wise insights directly on LeetCode problem pages — no LeetCode Premium required. It shows which companies have asked a given problem, how recently, and how frequently, with clean company badges and logos injected right below the problem title. A lightweight content script detects the active problem and pulls insights from a backend API, working seamlessly through LeetCode's SPA navigation.

ZKredit is a zero-knowledge proof based fair lending verification system — a bank proves that a loan decision came from its real, unmodified credit model, without ever revealing the model's weights or the applicant's private financial data. Real ZK-SNARK proofs (via EZKL and Halo2) are generated per application and verified both locally and on-chain, backed by a Solidity verifier and registry contract with a full applicant-facing auth flow (JWT + Google OAuth).

VelocityLLM is a dynamic/continuous batching scheduler engine for LLM inference serving, built on top of vLLM to prove that intelligent scheduling raises GPU throughput and SLA reliability over static batching. It enforces configurable latency SLAs via a predictive admission controller, adapts concurrency in real time with an AIMD feedback loop, and prevents starvation with a priority-aware queue — validated with real benchmark data showing up to 14.7% higher throughput and 100% SLA compliance under mixed load versus the static baseline.