An event-driven, high-throughput stream processing architecture for live financial fraud detection, combining a deterministic rule engine with a supervised XGBoost classifier that catches 98% of fraud at 82% precision (up from 28%/24% on an earlier unsupervised baseline). Kafka is partitioned by account ID for per-transaction ordering, paired with Redis-backed idempotency and per-message dead-letter-queue routing so a single malformed message can't take the system down. Dual Grafana dashboards — JVM health and custom AI-worker business metrics — sit on top of a Spring Boot + FastAPI pipeline with automated GitHub Actions CI/CD.

An ongoing multi-agent orchestrator designed to integrate modern AI with legacy enterprise systems, like COBOL mainframes, to solve process debt. Currently in early development, the architecture is kept lightweight by relying solely on Gemini API calls to coordinate specialized agents (Scout, Knowledge Bridge, Compliance).

A secure, local-first Retrieval-Augmented Generation (RAG) engine. Built with a dual-stack architecture to ingest enterprise documents, ensure multi-tenant data isolation, and stream context-aware AI responses in real-time.

A self-hostable RAG platform built from scratch to chat with local files privately without uploading to the cloud.

Bitwise is a desktop collaborative codespace built with Electron, React, and TypeScript. It provides a modern editor experience powered by Monaco, real-time collaboration via Yjs, and a connected backend for session management and persistence.

Finalist at the IIT Madras AUV Competition. Developed the core logic and navigation systems.

An end-to-end Machine Learning OCR pipeline utilizing CRAFT localization, custom Vertical Projection Profiling, and a fine-tuned Vision-Encoder-Decoder Transformer to digitize highly degraded 19th-century texts.