An AI-powered travel recommendation system built around a curated dataset of 100 offbeat Odisha locations unavailable on mainstream travel platforms. The system combines semantic vector search with LLM-driven itinerary generation and live flight/hotel data retrieval via the Amadeus API. Reached Top 50 nationally at Smart India Hackathon 2025.

Architected end-to-end from data curation to cloud deployment, resolving real infrastructure constraints and shipping a working product under competitive conditions. The offbeat Odisha dataset is the core differentiator — no equivalent exists in commercial travel APIs.

A multi-method medical image analysis pipeline comparing traditional and deep learning approaches for optic disc segmentation in retinal fundus images. Implemented and benchmarked morphological thresholding, K-Means clustering, DBSCAN, and a CNN U-Net architecture — achieving a best Dice score of 92.17% and IoU of 86.63%.

Designed to be methodology-honest — the comparative benchmark between classical CV methods and deep learning provides interpretability data clinicians can reason about, not just a black-box prediction.

A retrieval-augmented generation system that transforms uploaded study materials into an interactive AI tutor. Students upload course documents; the system builds a vector index and answers questions with source-grounded responses. Survived a full stack migration mid-development from Gemini Embeddings to HuggingFace all-MiniLM-L6-v2 after regional API access restrictions.

Built a complete RAG stack without relying on LangChain abstractions for the core retrieval logic — chunking, embedding, and similarity search are transparent and tunable.

End-to-end 3D dental image segmentation pipeline built as a take-home ML engineering assessment for Dobbe AI. Processes volumetric Cone Beam CT data through a SegResNet architecture using the MONAI framework, with interactive 3D visualization of segmentation outputs using Plotly.

Demonstrates the ability to rapidly onboard to an unfamiliar medical imaging domain, implement a state-of-the-art 3D segmentation architecture, and ship a complete, evaluated pipeline under time constraints.

A from-scratch Deep Packet Inspection engine built in Java over a structured 7-day development plan, targeting networking company recruitment. Implements packet capture, protocol dissection, and pattern-matching logic without relying on high-level DPI libraries.

Engineered with networking recruitment in mind — every architectural decision was made to demonstrate command of TCP/IP stack internals and Java systems programming.

An end-to-end sentiment analysis pipeline coupled with a production Power BI dashboard. Processes raw customer feedback through NLP preprocessing and classification, then surfaces insights via interactive DAX-powered visualizations optimized for business stakeholder consumption.

Bridges the gap between data science and business intelligence — a signal that engineering output should ultimately be readable by non-technical stakeholders.